Skip to content

AgentWeave v0.6.0 — Research Preview: Pre-Inference Routing for Tool-Rich Agents

Pre-release
Pre-release

Choose a tag to compare

View all tags
@sauravsingla sauravsingla released this 25 Aug 03:33
· 161 commits to main since this release
v0.6.0
24521be

AgentWeave v0.6.0 — Research Preview

AgentWeave is an open-source framework for deterministic pre-inference routing, policy-aware scope reduction, provenance, and reliability in tool-rich and multi-agent systems.

This release consolidates the current AgentWeave research and engineering foundation, including practical framework integrations, reproducible evaluation infrastructure, and a growing focus on failure analysis and recovery.

Highlights

  • Pre-inference routing — reduces the tool or agent set exposed to the model before inference.
  • Deterministic routing — supports reproducible tool and agent selection.
  • MCP integration — demonstrates routing over MCP-style tool descriptors.
  • LangGraph integration — applies AgentWeave routing before downstream graph execution.
  • AutoGen integration — supports pre-selection of agents and actions before orchestration.
  • Bring Your Own Model (BYOM) — enables alternative model integrations while preserving separation from the frozen research core.
  • Policy-aware scope reduction — provides a foundation for controlling which capabilities become model-visible.
  • Provenance and observability — supports analysis of routing decisions and downstream execution behavior.

Validation

The v0.6.0 release pipeline includes automated testing and package validation.

Current validation includes:

  • 83 automated tests passing
  • successful Python wheel build
  • successful source distribution build
  • successful twine check
  • release-tag and package-version consistency checks

Research evaluation

AgentWeave includes a frozen BFCL-derived routing-pressure study investigating whether reducing the model-visible tool space can affect downstream function-calling reliability.

These experiments are research evaluations and are not official BFCL leaderboard results.

Historical benchmark evidence remains frozen and is not retroactively modified by newer framework integrations or experimental extensions.

Failure analysis and recovery

Recent external research feedback has motivated a deeper evaluation phase focused on:

  • routing-stage failures;
  • function-selection failures;
  • multi-function composition failures;
  • execution correctness;
  • recovery after tool failure;
  • state changes during execution;
  • re-routing when previously pruned capabilities become necessary.

This work is tracked in Issue #29.

Evidence boundary

This release does not modify previously frozen benchmark results.

New integrations, BYOM functionality, recovery experiments, and future evaluation extensions remain separated from the historical experimental evidence to preserve reproducibility.

Status

Research Preview / Pre-release

AgentWeave is under active development. APIs, evaluation methodology, recovery behavior, and framework integrations may evolve before a production-stable release.

Feedback, reproducibility reports, issues, and research collaboration are welcome.

Assets 4
Loading