Who did this work?
Marcus · MIT license · in development since June 2025 · v0.3.8
Marcus is Gray Labs' flagship: an open-source multi-agent coordination system where a shared task board is the only channel. Agents never talk to each other. Each one pulls a task, does the work, and writes its progress, decisions, and artifacts back to the board. Anything that speaks MCP can sit at the board: Claude Code, Codex, Gemini CLI, an AG2 agent, or a person.
Because nothing can happen off the board, the board becomes the project's memory. The audit trail is not bolted on. It falls out of the architecture.
The four questions
Who did this work?
Why was it done this way?
What did it cost?
Does it actually work?
Companies are beginning to hand real software work to AI agents, and today these answers evaporate the moment each session ends. That is not a missing feature. It is a missing system of record.
On the record · v0.3.8
Register once, pull a task, receive the context of its dependencies, work, report, repeat. One agent per task.
Time-limited ownership with automatic recovery when an agent dies mid-task.
Every AI call writes one immutable row: agent, task, session, model, tokens. Prices are versioned by date, so cost is computed like a bank statement and history is never rewritten.
Every task event is logged with its task ID and can be replayed afterward in Cato, the companion dashboard.
Against chat coordination (AutoGen) and orchestrator coordination (LangGraph), the board won by roughly 3x at 9 tasks and 7x at 27, growing with project size. Public and reproducible in marcus-mini.
One published negative result: on small jobs, a single strong agent beat the fleet four minutes to thirty. Coordination has a cost, and I measured mine.
Landing now · ADR-0012
One fact drives the current migration: a single agent can hold far more work than it could a year ago. So decomposition moves up to the session level, work is cut at the seams between large chunks, and agents become long-lived sessions that pull work for hours. Sessions that large and that autonomous create three requirements.
A claim number on every assignment. Only the current number can complete a task, and a superseded attempt is preserved on the board, never silently discarded.
A silence timeout catches dead sessions the same way at every task size. A spending ceiling catches live ones stuck in a loop.
Done stops meaning the agent said so, and starts meaning tests ran and artifacts are attached.
Independent convergence
In June 2026 a Stanford group published DeLM (Mao and Mirhoseini): parallel agents sharing short verified notes through a common workspace, with a checker that rejects any note an agent cannot back with evidence. It scored up to 10.5 points higher on a standard coding benchmark at about half the cost, and removing the checker erased most of the gain. Independent validation of the substrate, and a warning: with the strongest base model, the advantage nearly vanished.
So the field's live question is the one Marcus is instrumented to answer: when do many agents beat one, as a function of model strength and task structure? The measurement program is the next phase, and collaborators are welcome.