personal-agent-harness

A lightweight personal AI agent harness with memory, task loops, tool adapters, and local-first safety rails.

This repo follows the current trend around small personal agents and agent harnesses such as Hermes Agent, nanobot, Goose, and OpenClaw-style local automation.

Goal

Create a minimal agent runtime that can run repeatable personal workflows without becoming a giant framework.

Core Loop

input -> plan -> choose tool -> execute -> observe -> reflect -> persist memory

Features To Build

• Local task runner for repeatable jobs.
• Simple JSON memory store.
• Tool adapter interface for shell, browser, files, and APIs.
• Approval gates for risky actions.
• Run logs that can be replayed and audited.

Repository Shape

src/
  agent/
  tools/
  memory/
  safety/
examples/
  daily-brief.json
  repo-maintainer.json
docs/
  architecture.md

Milestone 1

• Done: implement a CLI that accepts a task file.
• Done: add deterministic local tools so the harness is testable without paid APIs.
• Done: add file-read and shell-command tools behind approval gates.
• Done: store every run in runs/.

Run It

python -m personal_agent_harness.cli examples/daily-brief.json
python -m personal_agent_harness.cli examples/daily-brief.json --approve-shell

Test It

PYTHONPATH=src python -m pytest

Publish

This package is ready for GitLab Package Registry and PyPI releases. See RELEASE.md.

Trend Notes

• The strongest viral angle is "agent that grows with you."
• The useful product angle is "personal automation you can audit."
• Keep the first version boring, local, and reliable.