v0.13.3

Highlights

🛣️ Continuing the Path to v1.0

v0.13.3 brings a major Beta milestone — the cross-process Network — plus a new background_agent_tool, a Sandbox protocol, and evaluation improvements.

📍 Release Roadmap

---

🔬 Beta Framework (autogen.beta)

🌐 Cross-Process Network

The Network introduced in v0.13.0 ran in-process. v0.13.3 extends it across process boundaries — agents living in separate processes (and separate machines) can now participate in the same channels.

• 🔌 Cross-Process Data Plane – WsLink carries envelopes between hub and clients over WebSocket, with auth and delivery semantics.
• 🎛️ Cross-Process Control Plane – Frame-based RPC for the hub's control surface (registration, channel lifecycle, governance).

📔 Docs: Distributed Deployment

🤖 Background Agent Tool

• New background_agent_tool for running an agent in the background while the parent agent continues — useful for long-running work that shouldn't block the foreground conversation.

🏖️ Sandbox Protocol

• 🧰 Sandbox Protocol + LocalSandbox – New Sandbox protocol introduces a clean abstraction for sandboxed execution, with a LocalSandbox implementation that runs locally.

Other Beta Updates

• 🧪 Eval: Optional Reference Answer – Agent-as-a-judge and Pairwise judge can now exclude reference answers (autogen.beta.eval).
• 🔗 A2A – Create artifact before appending streamed text chunks (a2a-sdk 1.1.0); validate A2A protocol version during connection.
• 🧠 OpenAI Reasoning Tokens (fix) – Beta OpenAI client now maps reasoning tokens to thinking_tokens rather than cache_creation.
• 🌊 Stream – Guard Stream.get against a second matching event aborting the turn.

Bug Fixes

• 🔧 NoneType Crash – Prevent crash on missing tool result.

---

⚙️ Classic Framework

Bug Fixes

• 🔧 AutoPattern TerminateTarget – Prevent TerminateTarget from being silently ignored by AutoPattern.
• 🔧 OpenAI 2.41.0 Moderation Field – Sync the local ChatCompletion model with the new moderation field in openai 2.41.0.

What's Changed

• fix: prevent TerminateTarget from being silently ignored by AutoPattern by @Ricardo-M-L in #2606
• fix(beta/a2a): create artifact before appending streamed text chunks (a2a-sdk 1.1.0) by @marklysze in #2924
• chore(deps-dev): bump uv from 0.11.6 to 0.11.15 by @dependabot[bot] in #2925
• fix(beta): prevent NoneType crash on missing tool result by @marklysze in #2918
• fix(beta/openai): map reasoning tokens to thinking_tokens, not cache_creation by @vvlrff in #2928
• feat(beta/network): cross-process data plane (WsLink, auth, delivery) by @SirEntropy in #2914
• fix(beta/stream): guard Stream.get against second matching event aborting the turn by @vvlrff in #2929
• chore: fix typos in code and docstrings (INEGER, OepnAI, GrapStoreQueryResult, GraphRagCapbility) by @vvlrff in #2933
• fix(beta/a2a): validate A2A protocol version during connection by @vvlrff in #2935
• feat(beta/network): cross-process control plane via frame RPC by @SirEntropy in #2915
• feat(beta): add background_agent_tool by @vvlrff in #2900
• feat(beta/eval): Allow reference answer to be excluded for Agent-as-a-judge and Pairwise judge by @marklysze in #2936
• feat(beta): introduce Sandbox protocol and LocalSandbox by @vvlrff in #2879
• fix(oai): sync local ChatCompletion model with openai 2.41.0 moderation field by @marklysze in #2939
• fix(test): skip TinyFish tests when tinyfish is not installed by @marklysze in #2938
• fix(beta/network): make hub envelope ids strictly monotonic by @SirEntropy in #2941
• Version bump to 0.13.3 by @marklysze in #2940
• feat(beta/eval): export run_agent spans to an external backend by @marklysze in #2919

Full Changelog: v0.13.2...v0.13.3