Build out the agent runtime on top of the Agent API and tracing: tools
for composing agents, declarative tool policies, standard toolkits, and
pluggable state/artifact stores for resumable runs.
- Subagent tools: agent_tool() wraps an Agent as a callable tool and
propagates trace context to the nested run
- Tool policies: declarative AllowAllToolPolicy / DenyAllToolPolicy /
AllowToolsPolicy / RequireApprovalPolicy, ToolMetadata, and a tool()
decorator for attaching metadata
- Toolkits: standard file, shell, and git toolkits
- Persistence: StateStore (in-memory, file, Postgres) with compaction
records and StoredRunState for resumable runs, plus ArtifactStore
(in-memory, file) for large tool payloads
- Normalized model trace events (tracing/normalize) for consistent
model input/response capture
- tests for subagent tools, policies, toolkits, and state/artifact stores
Co-authored-by: Rohit <rohit.prasad15@gmail.com>
Devika