Changelog

Improvements

Faster agent runs with many tools: Tool schemas are now derived once and cached across runs, cutting per-run overhead for agents that carry large toolkits. (#9771)

Faster long conversations: Session history is loaded incrementally per turn, so response time stays flat as a conversation grows instead of scaling with its length. (#9775)

PubMed request timeout: PubmedTools now accepts a timeout and passes it to both NCBI E-utilities requests, so a stalled PubMed response cannot hold a tool call indefinitely. (#9463)

Bug Fixes

Strict-mode tool schemas without properties: Function.process_schema_for_strict no longer raises KeyError on schemas that omit properties (for example MCP server schemas registered verbatim). (#9578, fixes #9409)

Gemini tool-result media: Images and document blobs returned from tools are now nested inside FunctionResponse.parts for Gemini 3 and later models instead of being emitted as sibling inline_data parts; legacy models keep the previous representation. URI-backed response media on Vertex AI is only sent when the MIME type is supported. (#9647)

Location lookup on core-only installs: agno.utils.location used the undeclared requests package, so add_location_to_context=True raised ModuleNotFoundError without extras installed. It now uses httpx. (#9793, fixes #9772)

ScheduleManager cleanup: close() tolerates a partially constructed manager (for example after a failed deepcopy), so garbage collection no longer raises AttributeError on a missing _pool. (#9792)

Cookbooks

Added a deterministic side-effect approval flow example. (#9790)

Added DeepKeep AI Firewall guardrails example. (#9784)

Documented AgentOS team run states. (#9768)

Added a Peer Cash MCP agent example. (#9562)

Fixed stale cookbook paths left over from the directory renumber and in the MCP examples. (#9789, #9783)

What's Changed

• [fix] add PubMed request timeout by @Ghraven in #9463
• [cookbook] Add Peer Cash MCP agent by @ADWilkinson in #9562
• [cookbook] Document AgentOS team run states by @daleselaji-dev in #9768
• feat: export QueueConfig from agno.os by @ysolanky in #9776
• fix: correct stale 90_tools paths in MCP cookbook examples by @cc-9898 in #9783
• fix: cookbooks stale paths left by the directory renumber by @harshsinha03 in #9789
• [cookbook] Add DeepKeep AI Firewall guardrails by @gilarel in #9784
• fix: make ScheduleManager cleanup tolerate partial copies by @kausmeows in #9792
• [cookbook] Add deterministic side-effect approval flow by @daleselaji-dev in #9790
• fix: use httpx for location lookup instead of undeclared requests by @kausmeows in #9793
• fix: stop rebuilding session history on every conversation turn by @ashpreetbedi in #9775
• [fix] Nest Gemini tool-result media in function responses by @green3sf in #9647
• [fix] Tolerate a tool schema without properties in strict mode (#9409) by @Anai-Guo in #9578
• fix: cache tool schema derivation across runs by @ashpreetbedi in #9771
• fix: Drop no-op default assignments in search_user_memories by @darkdi in #9720
• chore: release 3.0.1 by @ysolanky in #9794

New Contributors

@ADWilkinson made their first contribution in #9562

@cc-9898 made their first contribution in #9783

@gilarel made their first contribution in #9784

@green3sf made their first contribution in #9647

@Anai-Guo made their first contribution in #9578

@darkdi made their first contribution in #9720

Full Changelog: v3.0.0...v3.0.1

Original source

⚠️ Breaking release. A database migration is required before v3.0 serves traffic. Read the v3 Migration Guide first, and see the full v3.0 Changelog for every change.

New Features

Tool result offloading: Agent(offload_tool_results=True) / Team(...) writes any tool result over 16,000 chars to AgentFS and leaves a short envelope (preview, size, result_id) in the message; the agent gets read_result / search_result (+ async) to fetch the rest. No model call on the write path. Tune via ResultStore(threshold_chars=..., ttl_seconds=...). (#9436, #9684)

Media offloading: media_storage=S3MediaStorage(bucket=...) on an Agent/Team/Workflow uploads images, audio, video and files to local disk, S3 or GCS before persistence; the row keeps a small MediaReference instead of base64 (a 113 KB JPEG drops from ~151,000 chars to 2,897). No schema change. (#9340) Docs

CodeMode: CodeMode(tools=[...]) swaps a wide tool schema for one programmable IPython kernel that persists across a session — the model writes Python and calls tools as awaitable handles, composing them (variables, loops, helpers) without round-tripping through the transcript.

FinanceTools: One unified finance toolkit with swappable data providers.

RampRouter model: New model provider for Ramp Router (router.com).

SuperGrok OAuth: Device-code auth for the xAI model.

AtomicMail Toolkit: Adds an AtomicMail toolkit (AtomicMailTools) that gives an Agno agent its own email inbox

Database

Runs get their own table: each run is a row in agno_runs with real columns (session_id, run_type, agent_id, team_id, workflow_id, user_id, parent_run_id, status, run_index) + JSON payload. Takes session write amplification from O(N²) to O(N) and removes the DynamoDB/Firestore item-size ceiling. session.get_messages(), get_chat_history(), db.get_session() and AgentOS session routes are unchanged (runs re-attach on read). (#8350)

Direct run APIs: db.get_run(), db.get_runs(session_id=..., status=..., limit=..., page=...), db.upsert_run(), db.delete_run(), db.delete_runs() (sync + async); db.get_session(runs_limit=N) and db.get_sessions(include_runs=False). (#8350)

One-line migration: MigrationManager(db).up() creates the runs store and copies legacy runs across, non-destructively and idempotently, on 12 sync + 4 async backends; un-migrated DBs still work (reads merge runs table with legacy blob). Schema versions tracked on every adapter. (#8350)

Typed stale-DB errors: MigrationRequiredError / SchemaMismatchError name both remedies; AgentOS carries error_id: "migration_required_error" in the JSON body. (#9669, #9631)

Three new tables: agno_runs, agno_jobs (durable background queue), agno_tool_results (offload index) — all auto-created.

Per-user isolation

Extends beyond sessions to metrics, schedules, evals, knowledge, components, entity memory, and 17 vector databases. Metrics aggregate per user per day; unowned components/knowledge are shared (readable by all, editable by admin) so enabling isolation doesn't 404 pre-isolation building blocks. (#8245, #8262)

AgentOS & Studio

Durable background execution: AgentOS(queue=QueueConfig(durable=True)) — accepted runs are committed rows that survive crashes/restarts/deploys, executed by any replica. Bounded concurrency (default 32, AGNO_BACKGROUND_MAX_CONCURRENCY), cancellable while queued, Idempotency-Key dedupe, 429 on full queue. Queue REST surface (GET /queue/jobs, .../{job_id}, POST .../requeue, GET /queue/stats). Redis is optional coordination, never truth. (#9079, #9504)

Studio 3.0 — governed catalog: create_* writes a DRAFT that serves nobody until publish_component; compare-and-set guards (typed 409s), tombstoned deletes, archive/restore, dependent-tracking. StudioTools returns a machine-readable envelope ({ok, status, data, error{...}, warnings}) across ~31 tools. (#9604)

Tools

MiniMax video generation tools. (#9529)

Toolkits now have a stable id used by AgentOS to reference tools.

Model Updates

Cerebras / CerebrasOpenAI default to gpt-oss-120b (was llama-4-scout-17b-16e-instruct). (#9244)

Gemini defaults updated to 3.7 Flash. (#9666)

Groq: replaced deprecated llama-3.3-70b-versatile with openai/gpt-oss-120b. (#9588)

OpenAI: reasoning_effort, reasoning_summary, service_tier, verbosity accept the full API value set (widened types; no call breaks).

Claude works with anthropic 1.0.0 (#9686): SDK-compat update for Claude models

⚠️ Breaking Changes

Every 2.x user must read this. A database migration is required before v3.0 serves traffic.

Storage & migrations

Runs are no longer a JSON blob in the sessions table — they live in agno_runs. Run MigrationManager(db).up() (or AgentOS POST /databases/all/migrate) before serving. The v2→v3 migration preserves the legacy runs column as a backup; reclaim it with db.cleanup_legacy_runs_column() (SQL) / db.cleanup_legacy_runs_field() (document/KV) after verifying.

Stale/unmigrated databases raise typed errors instead of silently misbehaving.

Pagination: page without a limit (or page < 1) now raises ValueError instead of returning unbounded/negative results.

AgentOS

JWT: secret_key removed from JWTMiddleware and authorization_config — use verification_keys (a list).

Metadata routes: GET /models removed (model data moved to GET /config under available_models); GET / returns a minimal landing response; GET /info is the single unauthenticated metadata endpoint.

MCP server config: AgentOS(enable_mcp_server=..., mcp_config=...) removed — pass a single mcp_server= instead.

Background execution requires a db on the component (returns 400 without one). External-framework agents (LangGraph, Claude, DSPy, etc.) stream inline for background=true and are not resumable.

Agents

Renamed params:

• enable_user_memories → update_memory_on_run
• search_session_history → search_past_sessions
• num_history_sessions → num_past_sessions_to_search
• num_past_session_runs → num_past_session_runs_in_search

reasoning=True removed — set reasoning_model=<native reasoning model> explicitly.

continue_run / acontinue_run: updated_tools removed — pass requirements (list of RunRequirement from the paused run output).

Culture feature removed: enable_agentic_culture, add_culture_to_context, CulturalKnowledge, culture tools, and the agno_culture table. Use Knowledge for shared cross-user info.

Teams & Workflows

The Workflow constructor is keyword-only: Workflow(name=..., steps=[...]). Team is unchanged — Team([agent_1, agent_2]) still works, though Team(members=[...]) is preferred.

Flat HITL kwargs removed on Step/Steps/Loop/Condition/Router (requires_confirmation, confirmation_message, on_reject, requires_user_input, user_input_message, user_input_schema, requires_output_review, output_review_message, requires_iteration_review, iteration_review_message, on_error, hitl_max_retries, hitl_timeout, on_timeout). Use human_review=HumanReview(...) (from agno.workflow.types); names unchanged except hitl_max_retries → max_retries, hitl_timeout → timeout.

Tools

MultiMCPTools deleted (along with allow_partial_failure) — use one MCPTools per server.

MCPToolbox: auth_tokens / auth_headers removed — use auth_token_getters.

DuckDuckGoTools.duckduckgo_search → web_search, duckduckgo_news → search_news (now built on WebSearchTools).

Flat Google tool modules deleted (agno.tools.gmail, googlesheets, googlecalendar, google_maps, google_drive, google_bigquery) — import from agno.tools.google.*.

Google tools: creds_path / auth_port → credentials_path / oauth_port; Sheets enable_read_sheet etc. → bare method names.

FileTools.check_escape → Toolkit._check_path (LocalFileSystemTools.check_escape unaffected).

SQLTools: enable_list_tables / enable_describe_table / enable_run_sql_query → bare method names.

Seltz: max_documents → max_results; the legacy SDK path is removed (seltz>=1.2.0 required).

StudioTool alias removed — use StudioTools.

BrightData.get_screenshot: unused output_path removed. PgVector.enable_prefix_matching removed (dead helper).

Knowledge & Vector DBs

Knowledge.add_content / add_content_async / add_contents_async removed → use insert() / ainsert() / ainsert_many().

GDriveContextProvider renamed → GoogleDriveContextProvider.

LanceDB: use_tantivy removed/ignored.

Searching a pre-v3 vector table with a user_id raises ValueError (directing you to the vector DB migration) instead of returning empty results.

Scheduler

update_schedule is now allow-listed (only name, description, method, endpoint, payload, cron_expr, timezone, timeout_seconds, max_retries, retry_delay_seconds, enabled, next_run_at, disabled_reason); any other key raises ValueError. Ownership/provenance/lock state are no longer writable via the generic path. New provenance columns added by the v3.0.0 migration.

The schedules unique key changes from name to (user_id, name). If duplicate schedule names exist across the same user, the migration aborts — deduplicate before migrating.

Evals

eval_id → run_id (#9739): eval classes no longer carry eval_id; every run gets its own run_id. store_result_in_file renames the eval_id parameter to run_id, the {eval_id} placeholder in file_path_to_save_results templates is no longer accepted (use {run_id}), and POST /eval-runs returns the id the row was actually stored under (run_id). Re-runs no longer overwrite each other.

Models & Learning

Mistral: mistralai v1 compatibility layer removed — agno[mistral] requires mistralai>=2.0.0 (and is back in the models extra).

agno.models.metrics module and the Metrics alias removed → import from agno.metrics (RunMetrics).

Model.classify_error removed → use ModelProviderError.classify(error).

Entity memory under namespace="user" is now isolated per user (row keys embed a user_id digest); pre-v3 rows re-keyed by the v3.0.0 migration (agno.learn.migrations.rekey_user_entity_learnings); EntityMemoryStore.delete/get require a keyword-only user_id in that namespace.

Removed learn aliases: MemoriesConfig → UserMemoryConfig; MemoriesStore → UserMemoryStore; Decision → DecisionLog.

Deprecated (still working)

knowledge_retriever(dependencies=...) → prefer run_context. A retriever whose signature still names dependencies keeps working via an explicit backward-compat branch; run_context wins when both are present.

Scopes: system:read / system:write → config:read / config:write. The old names remain valid aliases and existing tokens keep working.

RedisDB (vector DB) → RedisDb. Note RedisVectorDb is also still exported, to disambiguate from the agno.db.redis storage adapter.

Migration quick-reference

Docs: Step-by-step guide, database migration, and a paste-into-your-coding-agent prompt

import asyncio
from agno.db.migrations.manager import MigrationManager
# Step 1: run before serving v3.0 traffic (up() is async)
asyncio.run(MigrationManager(db).up())
# Step 2: VERIFY the runs landed before any cleanup
assert len(db.get_runs(limit=5)) > 0, "Migration copied nothing - do NOT clean up"
# Step 3 (optional, destructive): reclaim the legacy blob column.
# The migration preserves it as a backup, so force=True is required.
db.cleanup_legacy_runs_column(force=True) # SQL adapters
# db.cleanup_legacy_runs_field(force=True) # document / KV adapters

On AgentOS: POST /databases/all/migrate. Full details in the v3.0 changelog docs.

What's Changed

Update Cerebras defaults and cookbook models by @ryanl-cerebras in #9244
[fix] Repair stale cookbook links by @CRDong233 in #9590
fix: replace the deprecated Groq llama-3.3-70b-versatile with openai/gpt-oss-120b by @sannya-singal in #9588
[cookbook] Add OpenUI client example for AgentOS AG-UI by @vishxrad in #9605
feat: add emem cookbook example by @kumari-jaya in #9624
chore: update gemini defaults to use 3.7 flash by @markmcd in #9666
fix: docs clarify human-readable ID collision guarantees by @daleselaji-dev in #9665
fix: typos in code comments and docstrings by @feizhuzheng in #9597
[cookbook] Align Team HITL examples with cookbook standards by @Math1987 in #9671
cookbook: add emem entry to MCP cookbook README index by @kumari-jaya in #9636
fix: repair four imports that do not resolve in cookbooks by @tonydzi in #9498
feat: add MiniMax video generation tools by @octo-patch in #9529
feat: v3.0 by @kausmeows in #8210
feat: Release v3.0.0 by @kausmeows in #9755

New Contributors

@ryanl-cerebras made their first contribution in #9244
@CRDong233 made their first contribution in #9590
@vishxrad made their first contribution in #9605
@kumari-jaya made their first contribution in #9624
@daleselaji-dev made their first contribution in #9665
@feizhuzheng made their first contribution in #9597
@Math1987 made their first contribution in #9671
@tonydzi made their first contribution in #9498

Full Changelog: v2.9.0...v3.0.0

Original source

What's Changed

• test: stop three suites reading state they do not control by @ashpreetbedi in #9736
• fix: state the team leader's delegation rule as need, not as a comparison with itself by @ashpreetbedi in #9737
• fix: refuse dispatch cycles and bound dispatch depth in the Studio dispatch tools by @ashpreetbedi in #9745
• fix: end a no-task tasks run by answering, and forbid joined member ids on nested rosters by @ashpreetbedi in #9746
• fix: answer 422 with the validator's message when request validation fails by @ashpreetbedi in #9744
• fix: give every eval run its own run_id by @harshsinha03 in #9739
• feat: opt-in self_dispatch knob for the Studio dispatch guard by @ashpreetbedi in #9749
• fix: complete the workspace credential excludes, and stop Studio taking version 0 literally by @ashpreetbedi in #9752
• fix: hide inner sub-team member ids from the outer roster, and open with the leader's purpose by @ashpreetbedi in #9754
• fix: close a fail-open in the Studio edit guard, and make the bigger exclude list affordable by @ashpreetbedi in #9756
• fix: stop the knowledge config advertising readers this install cannot run by @ashpreetbedi in #9750
• chore: bump agno to 3.0.0a5 by @ashpreetbedi in #9757

Full Changelog: v3.0.0a4...v3.0.0a5

Original source

What's Changed

• feat: run SqliteDb in WAL journal mode and restore the durable benchmark row by @ashpreetbedi in #9707
• cookbook: install pydantic-ai-slim in the perf environment by @ashpreetbedi in #9706
• chore: v3.0 release readiness - put the database suites under CI and fix the gates that never ran by @ashpreetbedi in #9712
• feat: add MiniMax video generation tools by @octo-patch in #9529
• fix: stop test-a2a installing a lockfile that names an unpublished agnoctl by @ashpreetbedi in #9714
• fix: let expected_current_version guard a first publish by @ashpreetbedi in #9728
• fix: enforce Workspace exclude_patterns as an access boundary by @ashpreetbedi in #9730
• feat: make all telemetry calls fire-and-forget by @ashpreetbedi in #6458
• fix: let the team's own identity open its prompt, and stop the prompt describing a runtime we don't ship by @ashpreetbedi in #9731
• chore: bump agno to 3.0.0a4 and release agnoctl as 0.2.0a1 by @ashpreetbedi in #9735

Full Changelog: v3.0.0a3...v3.0.0a4

Original source

What's Changed

• fix: keep a workflow session's runs to its own runs by @harshsinha03 in #9634
• fix: point invalid-schema table errors at the migration path by @ysolanky in #9631
• fix: cache resolved tables in SQL adapters to avoid per-query existence checks by @kausmeows in #9623
• fix: install a usable Postgres driver with the postgres and dev extras by @ashpreetbedi in #9635
• fix: workflow event serialization survives cyclic object graphs by @ysolanky in #9641
• fix: file-generation import warnings fire only when PDF/DOCX is requested by @ashpreetbedi in #9640
• fix: AgentOSTools reports failed schedules, windows and refreshes correctly by @ashpreetbedi in #9242
• test: assert the archive race by what landed, not by which call raised by @ashpreetbedi in #9654
• fix: match a learnings record owner against the caller by value, not by type by @ashpreetbedi in #9655
• fix: isolate entity memory rows per user under namespace="user" by @ashpreetbedi in #9322
• feat: SuperGrok OAuth device-code auth for the xAI model by @Himanshu040604 in #9616
• fix: accept every reasoning effort and service tier the OpenAI API takes by @ashpreetbedi in #9658
• chore: update gemini defaults to use 3.7 flash by @markmcd in #9666
• fix: docs clarify human-readable ID collision guarantees by @daleselaji-dev in #9665
• fix: typos in code comments and docstrings by @feizhuzheng in #9597
• [cookbook] Align Team HITL examples with cookbook standards by @Math1987 in #9671
• cookbook: add emem entry to MCP cookbook README index by @kumari-jaya in #9636
• fix: repair four imports that do not resolve in cookbooks by @tonydzi in #9498
• feat: add MigrationRequiredError for stale database schemas by @ysolanky in #9669
• feat: CodeMode and result offloading for agents and teams (3.0 S1) by @ashpreetbedi in #9436
• feat: registry workflows, zero-config Studio, and six verified 3.0 regression fixes by @ashpreetbedi in #9639
• feat: CodeMode result_store handle, kernel fixes, and bounds by @ashpreetbedi in #9684
• feat: offload media from the database to local, S3, or GCS storage by @kausmeows in #9340
• fix: give the remaining-integration CI job a Postgres server by @ashpreetbedi in #9685
• fix: make the Claude models work with anthropic 1.0.0 by @ashpreetbedi in #9686
• refactor: import agno.agent in 147 ms instead of 233 ms by @ashpreetbedi in #9678
• refactor: drop the per-AgentOS Studio catalog binding, and keep stale-schema errors actionable by @ashpreetbedi in #9692
• refactor: cut avoidable per-run and per-chunk overhead on the hot paths by @ashpreetbedi in #9689
• fix: O(1) session lookups in InMemoryDb by keying storage on session_id by @ashpreetbedi in #9693
• chore: fix import sorting by @ashpreetbedi in #9697
• fix: make agno.workflow importable without fastapi installed by @ashpreetbedi in #9696
• feat: make LearningMachine the only Studio memory surface by @ashpreetbedi in #9695
• fix: invalidate the cached session when db is reassigned by @ashpreetbedi in #9699
• test: isolate global run-cancellation state between unit tests by @ashpreetbedi in #9702
• fix: expire cancel-before-start intents in the in-memory cancellation manager by @ashpreetbedi in #9701
• fix: cut per-turn history and session-save cost in long conversations by @ashpreetbedi in #9700
• cookbook: add the canonical performance benchmark suite with cross-framework comparisons by @ashpreetbedi in #9694
• chore: bump agno to 3.0.0a3 by @ashpreetbedi in #9704
• cookbook: refresh performance reference results after the long-conversation fixes by @ashpreetbedi in #9705

New Contributors

• @daleselaji-dev made their first contribution in #9665
• @feizhuzheng made their first contribution in #9597
• @Math1987 made their first contribution in #9671
• @tonydzi made their first contribution in #9498

Full Changelog: v3.0.0a2...v3.0.0a3

Original source

What's Changed

chore: remove deprecated enable_user_memories, search_session_history, num_history_sessions and num_past_session_runs params by @sannya-singal in #7834

feat: evals user-isolation by @harshsinha03 in #8262

chore: remove v3.0.0 evals migration by @harshsinha03 in #8448

refactor: drop flat HITL kwargs in favor of human_review=HumanReview(...) by @sannya-singal in #8354

feat: consolidate AgentOS metadata routes by @ysolanky in #7647

feat: add id to Toolkit by @ysolanky in #8724

feat: denormalize sessions table in db by @kausmeows in #8350

fix: derive runs table name from session_table to prevent silent run … by @kausmeows in #9388

feat: reliable background execution for AgentOS - bounded, observable, durable by @ysolanky in #9079

cookbook: make the at-most-once default discoverable where testers hit it by @ysolanky in #9492

fix: forward-port unified-continue gate removal to the agents door by @ysolanky in #9392

fix: fail-closed continue gate and reconciling sweep - the two P1 review findings by @ysolanky in #9404

fix: RBAC filtering covers DB-loaded teams and workflows on list endpoints by @ysolanky in #9505

fix: inline continue terminal sync survives client disconnect by @ysolanky in #9499

fix: Wave 2 reliability smalls - thirteen review findings, one commit each by @ysolanky in #9508

fix: thread-based heartbeats - lease renewal survives a starved event loop by @ysolanky in #9512

fix: Wave 3 reliability polish - fourteen review findings, one commit each by @ysolanky in #9527

feat: make Team and Workflow constructors keyword-only by @ysolanky in #9554

fix: external agents crash the background streaming route with raw ev… by @kausmeows in #9585

fix: sorting and pagination in queue by @mishramonalisha76 in #9504

feat: schedules, metrics, knowledge, vdb user-isolation by @kausmeows in #8245

Update Cerebras defaults and cookbook models by @ryanl-cerebras in #9244

[fix] Repair stale cookbook links by @CRDong233 in #9590

refactor: remove culture feature (experimental, v3.0 cleanup) by @Mustafa-Esoofally in #9515

refactor: remove reasoning=True shortcut, require explicit reasoning_model by @Mustafa-Esoofally in #8940

feat: improve agno create onboarding by @ashpreetbedi in #9600

fix: replace the deprecated Groq llama-3.3-70b-versatile with openai/gpt-oss-120b by @sannya-singal in #9588

refactor: remove deprecated API surface for v3.0 by @ysolanky in #9560

refactor: remove compat surface that never shipped a runtime deprecation notice by @ysolanky in #9593

refactor: remove mistralai v1 compatibility layer by @harshsinha03 in #9613

fix: repair garbled scheduler log messages by @ashpreetbedi in #9619

fix: case-sensitive knowledge test assertions fail on lowercase model output by @ashpreetbedi in #9618

feat: FinanceTools - one finance toolkit, swappable data providers by @ashpreetbedi in #9606

fix: gate stream param by stream_sub_agent_events flag by @Mustafa-Esoofally in #9115

[cookbook] Add OpenUI client example for AgentOS AG-UI by @vishxrad in #9605

feat: add RampRouter model for Ramp Router (router.com) by @ashpreetbedi in #9626

feat: add emem cookbook example by @kumari-jaya in #9624

feat: Studio 3.0 - the governed control plane for agents that build agents by @ashpreetbedi in #9604

New Contributors

@ryanl-cerebras made their first contribution in #9244

@CRDong233 made their first contribution in #9590

@vishxrad made their first contribution in #9605

@kumari-jaya made their first contribution in #9624

Full Changelog: v2.9.0...v3.0.0a2

Original source

Changelog

New Features

StudioRunnerTools: New identity-aware dispatch toolkit (agno.tools.studio_runner.StudioRunnerTools) that splits execution out of StudioTools. Any component (team lead, router) can mount it to discover and run Studio-built agents/teams/workflows without getting the Studio's create/edit/delete surface. run_* tools thread the caller's user_id into the sub-run so per-user state lands on the right person.

list_components name filter: Added a name filter to list_components.

Bug Fixes

A2A: Fixed the A2A stream client dropping Task-level metadata by breaking on status-update.

Workflow (WebSocket): Honor the selected workflow version over WebSocket.

Team HITL: Persist paused member runs so team HITL resume survives a session reload.

Rehydration: Preserve toolkit instructions on rehydration.

Framework annotations: Guard framework return annotations.

Behaviour & Breaking Changes

MCP tool tool_name override blocked (Security): MCP tool entrypoints no longer allow a call-time tool_name override. Previously a model could pass tool_name="delete_repo" to any MCP tool and the server would execute that tool while allow-lists, requires_confirmation, HITL approval, and logging all resolved from the declared name — bypassing any HITL/approval gate. The executed tool name is now closed over from tool.name. Model-supplied tool_name args are forwarded as ordinary arguments (not used to pick the tool), so tools that legitimately declare a tool_name parameter keep working.

Tool result caching now uses per-user keys (Security / behaviour change): With cache_results=True, the cache key now includes stable run-context identity (user_id, session_id), fixing a cross-user cache leak where a cached tool taking run_context (e.g. MemoryTools) served one user's result to another. run_id stays out of the key so caching remains useful across a user's runs. Existing cache behaviour changes — keys are composed differently, so prior cache hits won't line up the same way.

Rehydration now fails loudly (behaviour change): Deserializing a persisted component with unresolvable references used to silently degrade (an agent's tools became [], a team lost members, schemas/knowledge dropped) and then run. Unresolvable references now raise ComponentRehydrationError (an AgnoError, status_code=422) on strict paths. Strictness is a caller property: public from_dict/load default strict=False (round-trips keep working), but AgentOS lookups and every dispatch path (REST POST /runs, continue, MCP run tools, StudioRunner) default strict=True and now return a 422 naming the unresolvable piece instead of running a degraded component. Pinned member versions are also honored.

What's Changed

fix: MCP tool entrypoints must not allow call-time tool_name override by @ashpreetbedi in #9379

fix: preserve toolkit instructions on rehydration by @ashpreetbedi in #9395

fix: persist paused member runs so team HITL resume survives a session reload by @ashpreetbedi in #9396

feat: StudioRunnerTools - identity-aware dispatch toolkit for Studio components by @ashpreetbedi in #9371

fix: fail loudly on unresolvable rehydration references; honor pinned member versions by @ashpreetbedi in #9381

fix: guard framework return annotations; add name filter to list_components by @ashpreetbedi in #9382

fix: tool result caching - per-user keys, ToolResult round-trip, hooks on cache hits by @ashpreetbedi in #9380

fix: honor selected workflow version over WebSocket by @Ayush0054 in #9514

fix: A2A stream client drops Task-level metadata by breaking on status-update by @psinojiya in #9224

feat: Release 2.9.0 by @kausmeows in #9545

Full Changelog: v2.8.7...v2.9.0

Original source

Changelog

New Features

• AdvisorTools: Added AdvisorTools for asking advisor models for feedback.
• OpenRouteService Toolkit: Added OpenRouteService toolkit for accurate routing.
• StudioTools: Added component-aware schedule tools and history parameters.
• FileSystemTools: Allow overriding the FileSystemTools toolkit name.

Bug Fixes

• Persisted Components: Fixed toolkit-qualified tool rehydration for persisted components.
• Cohere: Stop dropping zero-valued sampling params (temperature, top_k, seed, frequency_penalty, presence_penalty now respect an explicit 0).
• HITL: Propagate top-level confirmation to tool_execution on requirement deserialization.
• Team.load: Fixed crash on SQLite with unexpected keyword argument label.
• Audio Tools: More robust audio tool-result handling (and switched the Smallest AI cookbook to Gemini).
• Dependencies: Exclude nltk 3.10.1, which breaks the unstructured import chain.

What's Changed

• fix: exclude nltk 3.10.1, which breaks the unstructured import chain by @sannya-singal in #9320
• fix: robust audio tool-result handling, switch Smallest AI cookbook to Gemini by @harshitajain165 in #9331
• fix: Team.load crashes on SQLite with unexpected keyword argument 'label' by @Himanshu040604 in #9337
• feat: add OpenRouteService toolkit for accurate routing by @ysolanky in #9287
• feat: allow overriding the FileSystemTools toolkit name by @ashpreetbedi in #9363
• fix: propagate top-level confirmation to tool_execution on requirement deserialization by @ashpreetbedi in #9351
• feat: component-aware schedule tools and history parameters in StudioTools by @ashpreetbedi in #9352
• chore: add scheduler extras to dev so CI exercises the schedule tests by @ashpreetbedi in #9366
• feat: add AdvisorTools for asking advisor models for feedback by @ysolanky in #7196
• fix: toolkit-qualified tool rehydration for persisted components by @ashpreetbedi in #9358
• [fix] Cohere: stop dropping zero-valued sampling params by @bharadwaj-pendyala in #9300
• fix: run sync scheduler DB calls off the event loop by @ashpreetbedi in #9370
• chore: Release v2.8.7 by @kausmeows in #9369

New Contributors

• @bharadwaj-pendyala made their first contribution in #9300

Full Changelog: v2.8.6...v2.8.7

Original source

Changelog

New Features

Smallest AI: Added SmallestTools, a text-to-speech toolkit for Smallest AI with text_to_speech (returns audio as a ToolResult artifact, optionally saved to disk) and get_voices. Supports the lightning_v3.1 and lightning_v3.1_pro models.

AgentOS: Added GET /metrics/refresh/status to observe background metrics refreshes:

Reports idle, running, completed or failed with started_at, finished_at and error, so clients can poll for completion instead of timing out silently. The state updates even when a refresh finishes without writing new data.

Available on the client as AgentOSClient.get_metrics_refresh_status()

OpenSearch: Added OpenSearch vector database support (agno.vectordb.opensearch) with vector, keyword and hybrid search in both sync and async variants, installable via the agno[opensearch] extra. Includes cookbook examples and a run_opensearch.sh script for local setup.

Improvements

Tools: Cached the Pydantic version lookup during tool wrapping. The package metadata was re-read on every wrap, which made repeated tool wrapping a hot path (100-wrap benchmark: 65.9 ms to 11.0 ms).

Bug Fixes

AgentOS: POST /metrics/refresh no longer blocks uvicorn workers on large datasets:

The sync calculate_metrics() call now runs in the threadpool, so other requests keep flowing during a refresh. Same 200 + metrics list response as before.

New opt-in ?background=true query param returns 202 immediately and runs the refresh as a background task. Background refreshes are single-flight per database.

GET /metrics with a sync BaseDb also moves its lazy refresh to the threadpool.

Encoding: Text-mode open() calls now pass explicit encoding="utf-8", so JSON cache and config files written on one platform stay readable on another (Windows defaults to cp1252).

What's Changed

feat: add Smallest AI text-to-speech toolkit by @harshitajain165 in #9015

docs: fix broken links in cookbook READMEs and CONTRIBUTING.md by @devdattatalele in #9195

[fix] Cache Pydantic version lookup during tool wrapping by @basnijholt in #9210

fix: add explicit encoding="utf-8" to text-mode open() calls by @Ghraven in #7984

fix: run metrics refresh in background to avoid blocking workers by @Mustafa-Esoofally in #9263

feat: support opensearch db by @anhphong22 in #3611

chore: release 2.8.6 by @ysolanky in #9271

New Contributors

@harshitajain165 made their first contribution in #9015

@devdattatalele made their first contribution in #9195

@anhphong22 made their first contribution in #3611

Full Changelog: v2.8.5...v2.8.6

Original source

New Features

AgentOS Tools: Added AgentOSTools to report on AgentOS usage, latency, failures, schedules, evals, components, and pending approvals.

Improvements

Traces: Added latency and error stats grouped by agent, team, workflow, or endpoint, plus tool and model call stats. Implemented for PostgresDb and SqliteDb.

Moonshot:

Default changed to kimi-k3.

Added use_thinking to toggle thinking mode.

Added file and video input support.

Bug Fixes

ClickHouse: Bind metadata keys and values as query parameters in delete_by_metadata.

Moonshot: Preserve reasoning_content across turns.

What's Changed

fix: eliminate SQL injection in ClickHouse delete_by_metadata by @VANDRANKI in #7883

fix: Update Moonshot for kimi k3 Add Moonshot thinking controls and examples by @RayST3 in #9057

feat: AgentOSTools read-only platform ops toolkit by @ashpreetbedi in #9185

chore: Release v2.8.5 by @harshsinha03 in #9187

Full Changelog: v2.8.4...v2.8.5

Original source

New Features

FileSystem (durable agent filesystem): A net-new state primitive — agents get a private, persistent filesystem with pluggable DB/local backends and fail-closed per-user namespace isolation. See cookbooks.

Improvements

Cookbooks: Rewrote the AgentOS cookbook (284 → 132 files, 24 lessons) and overhauled the agent quickstart for Gemini 3.6 Flash.

What's Changed

• cookbook: overhaul agent quickstart for Gemini 3.6 Flash by @ashpreetbedi in #9136
• cookbook: rewrite the AgentOS cookbook (284 files -> 132, 24 lessons) by @ashpreetbedi in #9153
• feat: FileSystem — durable agent filesystem by @ashpreetbedi in #9142
• chore: Release v2.8.2 by @kausmeows in #9154

Full Changelog: v2.8.1...v2.8.2

Original source

Changelog

New Features:

TwelveLabsTools: Added support for Marengo video embeddings. See docs.

Slack: respond_to_other_agents flag for peer-agent communication.

Improvements:

Learning Stores: extraction_tool_call_limit to prevent infinite loops.

Context Providers: stream_sub_agent_events supported across all providers.

Bug Fixes:

File Tools: list_files now exposes its optional directory parameter in the tool schema, so agents can list a specific subdirectory directly.

Teams:

Nested teams now retrieve their own team history.

Explicitly configured member IDs are now preserved as-is, so delegation to members with custom IDs works.

CSV Reader: read() now supports CSV file paths passed as strings.

Tools: Empty/missing argument invocation fix.

Breaking Changes

Scavio Tools: google_search now targets the Scavio Google v2 API for proper localization and paging. See Docs

What's Changed

cookbook: fix learning-zone casualties, add _00_quickstart, rewrite the runner by @ashpreetbedi in #9076

fix: resolve team member agents created via AgentOS UI/components API by @ysolanky in #9072

fix: errors when invoking tools with empty or missing arguments by @zouchanglin in #6303

[feat] Add Marengo video embeddings to TwelveLabsTools by @mohit-twelvelabs in #8973

fix: map Scavio Google search params to v2 (gl/hl/start) by @scavio-ai in #9021

fix: Expose FileTools directory parameter by @giri256 in #9078

fix: preserve explicit team member IDs by @khrnchn in #9011

[fix] Preserve nested team history in TeamSession.get_messages by @RushikeshGandhmal in #8968

fix: Support string paths in CSVReader by @sjsjsjjs534 in #8822

feat: add respond_to_other_agents flag for Slack peer-agent communication by @Mustafa-Esoofally in #8934

fix: add extraction_tool_call_limit to prevent infinite loops in learning stores by @Mustafa-Esoofally in #9077

fix: support stream_sub_agent_events in all context providers by @Mustafa-Esoofally in #9098

fix: add team_id filter to get_team_history for nested team support by @ProgrammerPlus1998 in #8956

chore: Release v2.8.1 by @kausmeows in #9135

New Contributors

@giri256 made their first contribution in #9078

@khrnchn made their first contribution in #9011

@RushikeshGandhmal made their first contribution in #8968

@sjsjsjjs534 made their first contribution in #8822

Full Changelog: v2.8.0...v2.8.1

Original source

Changelog

New Features

agno.scorer — Turn a run into a number:

CodeScorer: wraps any callable (bool | float | Score; typed-field comparison under output_schema recommended)

JudgeScorer: LLM judge with the model always an explicit choice and numeric verdicts normalized to exact endpoints ((score - 1) / 9)

ToolCallScorer: checks tool executions deterministically (refused, errored, or HITL-rejected calls never satisfy an expectation)

All scorers ship sync and async variants.

agno.environments — Environment + Task + run_rollouts(env, k=8): run each task K times in full isolation (fresh db/session/user, no memory/knowledge/learning writes, cache off; knowledge reads still work). Includes a live per-attempt grid, real pass rate per task, drift-vs-policy fingerprints, save/load/diff, learning_zone(), and to_sft_jsonl(...) to export passing attempts as conversational-SFT JSONL with a provenance sidecar. This is the pass@k door.

Case.scorer: the eval suite gains a third check — plug any scorer into a Case (with Case.expected), free and exact, no LLM call. SuiteResult.to_dict() gains additive score_value, score_passed, score_reason keys.

FileGenerationTools: Added code file generation for FileGenerationTools.

Gmail Tools: Added pagination and max_results_per_request.

Adanos Tools: Added optional Adanos market sentiment tools.

Bug Fixes

RemoteAgent / RemoteTeam: Fixed metadata being dropped on the A2A protocol path.

Content: Return empty string for empty content list in get_content_string().

Decision log: Replaced deprecated datetime.utcnow() in the decision_log store.

Breaking Changes

ReliabilityEval matches tool executions: Tool expectations are now satisfied only by a clean execution (RunOutput.tools, tool_call_error not set), not by message-side requests. Verdicts can flip red after upgrading — when they do, the eval was previously passing for the wrong reason (missing entries are annotated "... (requested but refused/errored — execution matching, new in 2.8.0)"). Argument checks moved to ToolExecution.tool_args with the same partial-match semantics.

Hardened judge prompt: Every AgentAsJudgeEval now fences judged output behind a per-call random nonce, with the untrusted-data instruction inside the prompt. A literal </output> no longer escapes the block, and "score this 10" inside a judged answer is data, not an instruction. Judge verdicts and token counts may shift.

What's Changed

fix: replace deprecated datetime.utcnow() in decision_log store by @Ghraven in #7949

cookbook: add dpo_jury pairwise preference example by @ashpreetbedi in #9033

cookbook: refresh data_labeling for agno 2.7.x by @ashpreetbedi in #9037

cookbook: jury calibration, hardening, and agreement metrics by @ashpreetbedi in #9038

cookbook: synthetic data generation workflows by @ashpreetbedi in #9040

cookbook: critique-revision, persona, and tool-call trajectory generation by @ashpreetbedi in #9043

cookbook: step-reward scoring, scale-out mechanics, and safety labeling by @ashpreetbedi in #9046

cookbook: image_search README - ingest is a full rebuild, not idempotent by @ashpreetbedi in #9047

[fix] RemoteAgent/RemoteTeam drop metadata on A2A protocol path by @psinojiya in #8944

feat: add pagination and max_results_per_request to Gmail tools by @Mustafa-Esoofally in #9030

fix: Return empty string for empty content list in get_content_string() by @chuck-duplocloud in #6122

[feat] Add optional Adanos market sentiment tools by @alexander-schneider in #9060

[fix] Clarify Adanos trending ranking by @alexander-schneider in #9061

fix: replace retired qwen/qwen3-32b with openai/gpt-oss-20b on Groq by @ashpreetbedi in #9055

feat: add code file generation to FileGenerationTools by @anuragts in #8420

feat: agno.scorer and the judge prompt fence by @ashpreetbedi in #9049

feat: rollout engine and Case.scorer seam by @ashpreetbedi in #9050

release: v2.8.0 by @ashpreetbedi in #9063

cookbook: expand environments into progressive verification suite by @ashpreetbedi in #9070

feat: Release v2.8.0 by @kausmeows in #9073

New Contributors

@psinojiya made their first contribution in #8944

@chuck-duplocloud made their first contribution in #6122

@alexander-schneider made their first contribution in #9060

Full Changelog: v2.7.4...v2.8.0

Original source