AgentComms

From the live deployment at agentcomms-cli.vercel.app

Problem

Coordinating several AI coding agents working on the same codebase at once is genuinely hard. The obvious first approach — have every agent read and write to one shared log file so the others can see what's happening — works for a day, then breaks down: the file grows without bound, agents start losing relevant context in the noise, and the models begin hallucinating state that was never actually true.

Audience

Developers and teams running multiple concurrent AI coding agents (Claude Code, Codex, OpenCode, and similar tools) against a shared codebase, who need real coordination instead of an ad hoc shared file.

My role

Sole builder. Started it during a six-week software development internship, after running into this exact coordination problem firsthand while trying to manage several agents by hand.

Key decisions

• Event-sourced rather than a single mutable log: every action (a message, a task claim, a handoff) is its own signed, ordered event instead of an edit to one shared blob of text — so state is reconstructed from a real history, not inferred from whatever text happens to still be in the file.
• A governed identity model: every agent (and the human owner) registers as a signed principal with a role and explicit scopes, so an agent can't act outside what it's actually been granted — including a passphrase-gated elevated key for the handful of actions (like granting another agent's scope) that should require a human in the loop.
• Structured task lifecycle instead of free-text status updates: tasks are offered, claimed, leased with a renewable expiry and a working-directory lock, and handed off explicitly — so two agents can't silently step on the same work at the same time.

Result

Grew from a small personal fix into what the project now describes as its own app, with supporting features beyond the original CLI. It's the actual coordination layer behind this site's own multi-agent development workflow — the mechanism this portfolio's own recent updates were coordinated through was AgentComms itself.

Constraints & honest scope

Built and used by one person so far, primarily for coordinating this developer's own concurrent-agent workflows; it hasn't been used by a team or published for wider adoption yet.