Agno 3.2.0 moves A2A to v1.0 and rebuilds Slack on Bolt
Updated Agno
Summary
Agno 3.2.0 (Nov 12, 2026) upgrades A2A to protocol v1.0, moves the Slack interface to Bolt, adds AgentOS homepage metrics and lets Super Agent use tools directly.
- Tool
- Agno
- Category
- Frameworks
- Date
- Open source
- Yes
- License
- Apache-2.0
- MCP
- Yes
- Pricing
- Not stated in the sources we cited.
Agno v3.2.0 was published on November 12, 2026 (20:47 Istanbul time), two weeks after 3.1.12. It is a minor version with real breaking changes: AgentOS and A2AClient move to the official a2a-sdk 1.x, the Slack interface is rewritten on Slack Bolt with stricter approval rules, and the OS Manager gets new defaults. On top of that come database-backed homepage metrics, stored ClaudeAgent and CodexAgent components, and a Super Agent that now does the work itself instead of handing requests to manager members.
Release facts
| Fact | Detail |
|---|---|
| Product | Agno, open-source framework and runtime (AgentOS) for agents |
| Version | v3.2.0 |
| Release date | November 12, 2026 (GitHub release) |
| Previous release | 3.1.12, October 29, 2026 |
| License | Apache-2.0, per the GitHub API |
| Python | >=3.9,<4, per PyPI |
| Install | pip install -U agno==3.2.0 |
Check before upgrading
- Upgrade A2A server and client together. AgentOS and
A2AClientnow run ona2a-sdk1.x, and thea2aextra requiresa2a-sdk[http-server]>=1.2.2,<2. The/v1/message:*and/v1/tasks:*routes are gone; call each entity's own A2A endpoint. Stream events and artifacts follow the v1.0 shape. Standard v0.3 clients still connect through a compatibility wire, which is on by default. - Slack runs on Bolt now. Run
pip install -U "agno[slack]", which needsslack_bolt>=1.29.0,<2andslack_sdk>=3.43.0,<4. Only the person who started a run can Approve or Deny it, unless they are listed inapprovers=or mapped throughusers=. Internal helpers such asSlackEventHandler,HITLHandlerandverify_slack_signaturewere removed. - AgentOS can refuse to start. It stops at startup if Slack serves
agent_os.managerunderAgentOS(authorization=...)withoutusers=, ifOSManager(root=...)is combined with your own tools namedread_file,list_filesorsearch_content, or ifSlack(...)has no token or signing secret. - OS Manager defaults changed.
tool_call_limitgoes from 8 to 24 andnum_history_runsfrom 7 to 5. Set them explicitly if you tuned costs around the old values.
What's new
- Super Agent uses tools directly. It lists and runs agents, teams and workflows, reads platform activity and schedules, and uses Company, docs, source-file and custom OS Manager tools. Platform Builder is its only member, and every tool keeps its own permission check.
- Company OS. Routines run on a restricted Mini Agent that cannot chat and is hidden from discovery, MCP and A2A. Tasks expose
last_runandallowed_transitions, can be started directly, and Super Agent can start an accepted Task on the durable queue without blocking the conversation. Published Studio agents, teams and workflows can execute Tasks and Routines; drafts cannot. - Homepage metrics. New tables back daily sessions, runs, tokens, users, models, cost, duration, time to first token and model latency under
/os/metrics/*. Refreshes run in the background, rebuild only changed days and support ETags. Fresh installs no longer create the oldagno_metricstable. - Stored external agents.
ClaudeAgentandCodexAgentcan be saved, versioned, listed and run through the components API. Settings that cannot be serialized now fail loudly instead of being dropped. - A2A tasks in the database. Send, stream, get, cancel, list and subscribe all work on stored tasks, human-in-the-loop pauses can continue, and Agent Cards advertise both v1.0 and v0.3.
- Slack agent view. Streaming replies with task cards, a Stop button that cancels Agents, Teams, Workflows and external agents, long runs that continue in a new message after nine minutes, and Claude or Codex approvals answered from Slack.
- Paginated session runs.
GET /sessions/{session_id}/runstakeslimit,before_run_indexandafter_run_index. Calls withoutlimitstill return a plain list.
Notable fixes
- Run and session token totals no longer double count and now include failed or cancelled external runs.
- Redis is optional for sandboxed agents on the durable queue; Postgres plus a shared checkpoint store handles cancellation and resume across replicas.
- Mistral streaming calls send the output schema again, and Hugging Face streams keep every tool call instead of only the first.
- Gemini Developer API features default to
v1beta, andgeneration_configvalues are no longer overwritten.
Questions
Will Agno 3.2.0 break my deployment?
It can if you use A2A routes under /v1/message:* or /v1/tasks:*, the Slack interface, or OSManager(root=...) with custom file tools. Check those first, then review the new OS Manager defaults.
Do old A2A clients still work?
Standard v0.3 clients do, through the compatibility wire that is on by default. Agno's own A2AClient should be upgraded together with the server.
Who can approve a run in Slack now?
The person who started it, plus anyone named in approvers= or mapped to the same AgentOS user with users=.
More on the Agno hub, the previous Agno 3.1.12 launch and the frameworks category. Weighing a switch? Read Agno alternatives, Agno vs CrewAI and Agno vs LangGraph.
Alternatives
| Tool | License | Pricing note |
|---|---|---|
AgnoThis launchAgno is an orchestration and service layer for agents, exposing them through a REST API and an MCP server. | Apache-2.0 | Not stated |
CrewAICrewAI is a framework for orchestrating role-playing, autonomous AI agents that work together. | MIT | Not stated |
LangGraphLangGraph is an open-source orchestration runtime for building stateful, resilient agents. | MIT | Not stated |
OpenAI AgentsOpenAI's agent harness, shipped as a Python SDK update in April 2026 and as the hosted Agents API in September 2026. | Open source | SDK and Agents API use standard token and tool pricing. The Agents API has no additional fee. |
More from Agno
Agno 3.1.12 brings native Claude and Codex approvals into AgentOS
AgnoNative permission requests can pause persisted runs and resume the same execution after review.
Category: Frameworks
Pricing: Not stated
Agno 3.1.11 puts Claude and Codex harnesses inside teams
AgnoHarness agents can join teams and workflow steps, while new session-search endpoints support nested filters.
Category: Frameworks
Pricing: Not stated
Agno 3.1.5–3.1.6 adds repository sandboxes and session-aware evals
AgnoAgno adds repository-pinned sandboxes, S3 checkpoint storage, on-demand schedules and evaluation cases with session context and file inputs.
Category: Frameworks
Pricing: Not stated
Agno v3.1.4 adds sandboxed Claude and Codex agents
AgnoAgno v3.1.4, released October 16, 2026, adds Docker and AgentCore sandboxes for ClaudeAgent and CodexAgent, with checkpoint-based session recovery.
Category: Frameworks
Pricing: Open source
Agno v3.1.3 adds background runs to Claude and Codex agents
AgnoAgno v3.1.3, released October 12, 2026, adds background runs, retries, metrics and structured output to ClaudeAgent and CodexAgent external agents.
Category: Frameworks
Pricing: Open source
Questions
What is Agno?
According to its documentation, Agno helps companies run agents as a service. You build agents with the Agno SDK, Claude Agent SDK, Codex SDK or another supported framework. Agno then runs them as a service with durable execution and multi-replica scaling, exposed through a REST API and an MCP server.
Is Agno open source?
The GitHub API reports the Apache-2.0 license for the agno-agi/agno repository. The cited sources do not give pricing for any hosted offering.
Does Agno support MCP?
Yes. The documentation says Agno exposes agents through a REST API and an MCP server, which lets you make a product available in AI apps such as Claude and ChatGPT. The docs can also be given to coding agents through MCP.
Filed on Agent Radar. Claims in this note are limited to the sources above.