Shared Task Board Coordination
Multi-agent work can be coordinated through a durable, dependency-aware task board in which every cross-role decision is a card comment, a dispatcher promotes cards when their dependencies finish, and the coordinator has no tools to do the work itself.
Evidence: weak
When several agents share a task, the coordination itself becomes a failure point. Decisions made in one agent's context never reach another, work starts before what it depends on is finished, and afterwards nobody can reconstruct why the result looks as it does. The MAST taxonomy puts inter-agent misalignment at about a third of observed multi-agent failures.
Meta's multi-agent recipe answers this with a shared Kanban board, running on Hermes Agent. Four agent profiles (product manager, backend, frontend and go-to-market) work from one board stored in SQLite and dispatched from inside the Hermes gateway. Cards move between ready, running, blocked and done; parent and child links hold a card in ready until its dependencies are done, so sequencing comes from dependencies rather than polling, much like a task in a DAG workflow waiting on its upstream tasks. The product-manager profile is the only arbiter and deliberately has no terminal or code-execution tools. A specialist that needs input blocks its card, the manager answers in a card comment and unblocks it, and rework goes back as a new card with file-level acceptance criteria. The recipe's rule is that every cross-role decision is a comment on a card, "never a back-channel chat", so the run can be audited and replayed later. It recommends a single coding agent instead for one-role tasks such as a bug fix.
Compared with free-form agent chat, the board is a structured message channel of the kind inter-agent coordination protocols argue for, and it plays the role of a session log outside the context window for the team. It keeps the coordinator asymmetry of single-threaded writes while allowing parallel specialists. A board with decisions as comments is also a natural home for a full-trace audit of what a team of agents did and why.
The recipe also documents its own failure modes, which generalise to any dispatcher-driven board: cards sit in ready forever if the dispatcher is down, misspelt agent names fail silently, rate limits or context overruns crash a specialist until someone unblocks the card by hand, and a clarifying question that times out (after 600 seconds in Hermes) blocks the run. A stalled dispatcher or a card stuck in ready is therefore a failure signature worth monitoring for. The evidence is one worked vendor example without results, so the note is weak.
Sources
- Meta Model API cookbook, Use Cases (Hermes Agent with Muse Spark), undated (retrieved 3 Oct 2026) — https://dev.meta.ai/docs/cookbook/multi-agent-orchestration
- Cemri, Pan, Yang et al., arXiv 2503.13657, NeurIPS 2025 Datasets and Benchmarks (MAST) — https://arxiv.org/abs/2503.13657