AG2 now targets the Anthropic 1.x and OpenAI 3.x SDKs, TealTiger governance gains two more policies, and ACP servers can resume sessions.
What's Changed
Provider SDK support
Two provider SDKs released new majors, and AG2 has been adapted to both.
- Anthropic SDK 1.x — AG2 adapts to the breaking changes in the Anthropic SDK 1.x line, with Anthropic model name references updated to match.
- OpenAI SDK 3.x — Support for the OpenAI SDK 3.x line, including hosted shell and MCP events.
- AnthropicBashTool — New built-in tool, with improved event handling for Anthropic.
If you pin either SDK yourself, check your bound before upgrading.
Extension: TealTiger governance — two new policies
The governance middleware gains two policies, on top of the tool allowlist, PII, secret, cost and prompt injection checks already in the set:
tool_blocklist— Deny specific tools by pattern, the inverse of the allowlist, for cases where blocking a known-bad set is clearer than enumerating everything permitted.arg_validation— Validate tool arguments before the call runs.
Contributed and maintained by @nagasatish007, Documentation.
ACP session resume
session/load— An AG2 Agent served over ACP can now resume a prior session, rather than every connection starting fresh. Documentation.
Evaluation and telemetry
- Trace conversion is public API —
readable_spans_to_traceandspans_to_traceare now exported, so you can turn spans into traces in your own tooling rather than reaching into internals. - Structured tool results on spans — Tool spans now record structured and multi-part tool results, instead of flattening them.
- Tasks without traces are counted — Suite tasks that produce no trace are counted rather than dropped, so a suite total reflects what actually ran.
MCP, Sub-agents, and AG-UI
- MCP namespaced toolkit tools — Toolkit tools can be namespaced, so tools from different toolkits no longer collide on name.
- Bounded subagent fan-out — Concurrent task fan-out to subagents can be bounded, giving you a ceiling on how many run at once.
- AG-UI token usage — Stream events now carry token usage reporting.
Fixes and documentation
- Middleware — Retries stop once streaming chunks have been emitted, so a retry can no longer duplicate output the caller has already seen.
- Arize — Telemetry and evaluation options clarified.
New Contributors
- @Whxuan0701 made their first contribution in #3180
- @arizedatngo made their first contribution in #3208
- @junnhwan made their first contribution in #3216
Full Changelog: v1.0.3...v1.0.4