Skip to content

Roadmap

Bayesian-Agent v0.5 is a native-first early release. The core package is usable for evidence ingestion, belief updates, context rendering, repair planning, result summarization, and first-party benchmark execution.

The roadmap is organized around the project's main advantage: Bayesian-Agent should support full self-evolution from scratch, incremental repair for existing agents, and cross-harness adaptation instead of becoming another isolated agent framework.

Completed

  • Refactored the GenericAgent prototype into a standalone package core.
  • Defined a common trace schema for agent runs.
  • Implemented the Bayesian Skill registry.
  • Implemented full self-evolving primitives.
  • Implemented incremental repair utilities.
  • Added a GenericAgent optional adapter boundary without vendoring GenericAgent.
  • Added the first-party Bayesian-Agent native harness.
  • Added optional mini-swe-agent and Claude Code backend boundaries.
  • Released experiment result artifacts.
  • Added bilingual README files.
  • Added MkDocs documentation and GitHub Pages deployment.

Next

  • Harden executable benchmark runners for native and external backends.
  • Add richer rewrite policies and adapter examples.
  • Add adapters for more agent harnesses after the current compatibility boundaries stabilize.
  • Add more examples for project-specific failure-mode taxonomies.
  • Add documentation for operating Bayesian-Agent in a continuous evaluation pipeline.

Bayesian Algorithm Direction

The current v0.5 implementation defaults to a per-Skill Bayesian Evidence Model with a categorical likelihood backend and keeps Beta-Bernoulli as an optional compatibility backend.

Future releases will move toward broader Bayesian reasoning:

  • Skill hypothesis inference: compare, combine, and specialize competing Skill/SOP hypotheses with posterior evidence.
  • Bayesian Networks and graphical structure: model dependencies among tasks, contexts, tools, failure modes, and Skills.
  • Uncertainty-aware online decisions: choose selection, rewrite, rerun, retire, and transfer actions under uncertainty as harnesses, models, and tasks change.

Non-Goals for v0.5

  • Bayesian-Agent does not train or fine-tune base models.
  • Bayesian-Agent does not replace GenericAgent.
  • Bayesian-Agent does not copy or vendor GenericAgent.
  • Bayesian-Agent is not limited to GenericAgent.
  • MinimalAgent adapter support is not included yet.