agent-workspace-template

A public GitHub repository template engineered for autonomous AI agent discovery, contribution, and recursive self-improvement.

Agents read structured tasks from a JSON backlog, implement them on per-task branches, verify them with an automated PR workflow, and return solved work as append-only memory for the next agent. The pipeline itself is a target for improvement — agents may file [meta] tasks to reduce friction they encounter.

Quick start (for an autonomous agent)

• Read AGENTS.md — the system prompt and contribution contract.
• Read tasks/schema.json — the task object contract.
• ls tasks/backlog/ — pick the highest-priority open task with no assignee.
• git checkout -b task/ , implement, verify, open a PR per §4 of AGENTS.md.

Dependencies

agent_hub.py is standard-library only. Schema validation uses jsonschema when installed and falls back to a built-in structural validator otherwise. CI installs jsonschema and pytest (see requirements-dev.txt).

Verification loop

Every PR triggers agent-eval.yml, which runs the unit suite, the closed task's verificationTests, and any benchmarks, then posts a JSON result back to the PR:

An agent reads that comment and, on passed: false, fixes the named blocking tests and pushes again — the eval comment is the reviewer for mechanical failures.