virtual-agent-org

A network of Claude agents that helps run the Accord Project: triaging and reviewing pull requests, measuring community health, and turning Tech Working Group recordings into summaries. The TSC chair instructs one agent, Ada, the Chief of Staff, from Telegram or Discord. Ada delegates to specialist agents, and every critical action (merging, publishing, posting, messaging someone) waits in a decision queue for the chair's approval.

Status: Phase 1, an experiment. Five agents (Ada, Mira, Vera, Otto, Sage) work on Concerto, Template Engine and Template Playground. Agent-drafted messages are labelled as such. A sixth agent, Mandla, builds changes to this repo's own specs and tools that the chair has approved, as pull requests he reviews.

How it fits together

1. Each agent runs on Claude Managed Agents with its job spec from agents/, its tools, and its own credentials vault and memory.
2. Agents call custom tools (request_action, assign_task, report_status, …). Managed Agents notifies the gateway, a Cloudflare Worker, which runs the tool.
3. request_action queues a decision and returns its id at once, so the agent keeps working. The chair approves or rejects it from a card in Telegram or Discord.
4. On approval the gateway carries out the action with its own credentials and logs it.

ARCHITECTURE.md explains the design, the safety rules and how the network is governed; docs/decisions.md records why.

Layout

Path	What it is
ARCHITECTURE.md, docs/	Design, decision log, and setup guide for running your own copy
agents/	One job spec per agent, plus _protocol.md, which every agent follows
config/network.yaml	Phase, agents per phase, repos in scope, budgets
config/permissions.yaml	Action kinds and their risk, and what each agent may use and request
config/models.yaml	Model tiers, and effort per agent and per task; specs name a tier, never a model
config/schedules.yaml	Recurring work: check-ins, triage, reviews, scorecards
config/people.yaml	People who can talk to the network, their roles and GitHub logins
config/mcp_servers.yaml	External tool servers and where each agent's token comes from
context/accord.md	Shared, read-only project context for every agent
evaluation/	Weekly evaluation rubric and scorecard schema
gateway/	The Cloudflare Worker: bots, decision queue, custom tools, audit log
scripts/	Config validation, deploys, and helpers such as repo:add

Talking to the network

Plain messages to the Telegram bot go to Ada. On Discord, use /ada <text> in the server or in a DM with the bot. Commands: /status, /queue, /ask <agent> <text>, /pause <agent>, /resume <agent>, /stop <agent|all>, /revoke <agent> confirm. People listed in config/people.yaml can ask questions and see status and the queue; only the chair can instruct, approve or stop agents.

Development

Requires Node.js 22 or later.

npm ci && npm ci --prefix gateway
npm test                     # config validation, script tests, type-check, gateway tests
npm run compile              # regenerate gateway/src/generated/network.json after a config change
npm run deploy:agents:dry    # what an agent deploy would change
npm run repo:add owner/name  # check a repo is ready, then add it to the network

Deploying needs the project's Anthropic, Cloudflare and GitHub credentials, so it is done by the maintainers: merging to main starts a deploy, which runs once the owner approves it. To run your own network for another project, see docs/setup.md.

Contributing

Contributions are welcome: see CONTRIBUTING.md. Changes to agent specs, permissions and models are proposed as pull requests and merged by the chair. To report a security issue, see SECURITY.md.

License

Apache-2.0. See LICENSE.