Multi Agents

CLI-first multi-agent workflows for work operations. The MVP takes local files as input, runs specialist agents, passes outputs through review, and saves Markdown reports plus JSON trace summaries.

Install

npm install
cp .env.example .env

Set OPENAI_API_KEY in .env. Override OPENAI_MODEL when needed.

Run

Start with a template when using real work material:

npm run templates
npm run config:check

npm run workflow -- research-report examples/inputs/research-report.md
npm run workflow -- ops-weekly examples/inputs/weekly.md
npm run workflow -- meeting-actions examples/inputs/meeting.md
npm run workflow -- content-ops examples/inputs/content.md

For a no-key local preview:

npm run workflow:mock -- research-report examples/inputs/research-report.md

Each run writes:

• runs/reports/<run-id>.md
• runs/traces/<run-id>.json

Record whether a run helped:

npm run evaluate -- runs/traces/<run-id>.json --before-min 90 --after-min 35 --adopted partial --error "missed owner" --rework "clarify metric source"
npm run review -- --out runs

Compare two runs when checking prompt or workflow config changes:

npm run compare -- runs/traces/<base-run-id>.json runs/traces/<candidate-run-id>.json

Evaluation records are written to:

• runs/evaluations/<run-id>.json
• runs/evaluations/index.jsonl

Workflows

Workflow definitions live in config/workflows/*.json and are validated with:

npm run config:check

Set AGENT_WORKFLOWS_CONFIG_DIR to point the CLI at another workflow config directory.

Each workflow config includes a required semantic version. Run traces record the workflow version, a short hash of the raw config JSON, and the centralized agent prompt version so later reviews can explain why two runs differed.

• research-report: parallel research notes, fact check, contrarian review, final synthesis.
• ops-weekly: progress collection, metrics synthesis, blockers, weekly report.
• meeting-actions: transcript/materials to action items, priorities, risks.
• content-ops: topic framing, research, draft review, fact and policy check.

Safety Boundary

The system never sends emails, publishes content, deletes files, opens pull requests, deploys, or commits on its own. High-impact actions are listed under Human Approval Required for manual review.

Development

npm run typecheck
npm test
npm run build

After changing workflow config, run:

npm run config:check
npm run workflow:mock -- ops-weekly examples/inputs/weekly.md

Two-Week Use Loop

Use ops-weekly and meeting-actions first. After every real run, record time saved, errors, rework, and adoption with npm run evaluate. Review the aggregate with npm run review -- --out runs at least weekly. When prompt or config changes are being tested, run the same input before and after the change and compare traces with npm run compare -- <base-trace> <candidate-trace>. Stabilize only workflows that are used three or more times per week or clearly save recurring review time.

Expansion Ideas

Use docs/roadmap.md and templates/evaluation/workflow-idea.md before adding new workflows. Promote ideas that are frequent, file-based, review-heavy, and safe to keep behind human approval.

See docs/workflows.md for the workflow config checklist.