The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation
Tyler Hunt, Donald Nguyen, Gurbinder Gill
Timely, well-argued systems design with candid limitations and reusable concepts, but evidence is small synthetic pilots and the headline cross-vendor claim is untested
Abstract
Harnesses for large language models have converged on a single-threaded shape: one agent, one history, one machine, one sitting. That shape fails work that spans many agents, outlives any one model call, or waits on systems that are not agents. We describe the Agent Board, the coordination substrate in the Corvic harness. The board is a durable, shared table of semi-semantic tasks: a few exact fields the system routes on, and a free-form body that only language models read. Agents post tasks, watch tasks, and block on them, ending their turn and holding no process until a matching post wakes them. A system agent, the Board Keeper, acts as the interrupt handler for the board: it carries events from non-agent parts of the harness onto the board, matches them to waiting tasks by reading their prose, and allocates new agents for tasks that name a kind of agent rather than an instance. Loadable coordination policy kernels give statuses and roles meaning for a given pattern of work, without the board enforcing them. The design mirrors the kanban board a human team puts on a wall. We argue it is what persistent agent products such as Meta Muse, OpenAI dots and SpaceXAI Grok Bot need once an agent must wait, hand off work, or work with agents from another vendor. We view the board as an operating-system scheduler for agents and state where the analogy breaks. We separate what the mechanism guarantees from what it leaves to convention, and report a trace from a live board and three small simulated pilots. In the pilots, the wait cost of the board was flat in wait length while naive polling grew about 165-fold, but a plain external watcher was cheaper than the board. The Board Keeper matched prose waiters with no false wakes but missed 3.6 percent of real wakes.
AI Impact Assessments
(1 model)Core Contribution
The paper presents the Agent Board, the coordination layer in Corvic's LLM harness. It is a durable shared table of "semi-semantic" tasks. A few exact fields (id, topic, status, assignees) are routed on by the system, and a free-form prose body is read only by LLMs. Its main primitive is blocking: an agent posts or watches a task, ends its turn, and holds no process until a matching post wakes it. A system agent, the Board Keeper, acts as interrupt handler. It turns non-agent events (pipelines, quotas) into board posts, matches them to waiters by reading their prose, and allocates agents by kind rather than by instance. Loadable "coordination policy kernels" (kanban, delegation, swarm, escalation, pipeline-wait) supply protocol as instructions rather than enforcement. The paper also gives an OS-scheduler analogy with its break points, an explicit list of guarantees (G1–G6) against what is left to convention, a trace from a live board, three small simulated pilots, and a benchmark proposal (WaitBench).
The most original ideas are the prose-valued wait condition matched by an LLM, the level-triggered wake-on-watch rule that closes the lost-wakeup race, and the argument that a durable board, not a wire protocol, is the right rendezvous point for cross-vendor agents. Much of the substrate is old (blackboards, Linda tuple spaces, Contract Net, condition variables, Kubernetes reconciliation). The novelty lies mostly in recombining these with an LLM as the matcher and in the framing.
Methodological Rigor
This is primarily a systems and position paper, and its self-assessment is unusually honest. The pilots are small simulations with one small model and synthetic tasks, none run on the production board.
The weaknesses are no production measurements, a single model, small n, no tests of scale or adversarial conditions, no deadlock detection, no mutual exclusion, and no benchmark release. The claims are generally well hedged against the evidence.
Potential Impact
If the board concept proves right, it could influence how persistent agents coordinate across vendors, a live concern as always-on agent products proliferate. The conceptual contributions are reusable ideas:
WaitBench's "hold the model fixed, vary the harness" framing is a sensible gap-filler, but it is only a proposal with no promised release, which limits its impact. Practitioners could adopt the pattern readily, since it is implementable on any database plus a notification service. Academic uptake is limited by the lack of rigorous evaluation. The paper is also tied to a commercial product (Corvic), and the strongest cross-vendor claims are speculative ("not an integration any vendor ships today").
Timeliness
The paper is highly timely. Long-running, persistent, multi-agent systems are the emerging bottleneck, and the paper engages current products (Muse, dots, Grok Bot) and benchmarks (Gaia2, METR, Handoff Debt). Interop discussion centers on transport (MCP, A2A), and the paper argues effectively that waiting and records are the missing layer. Risk: the area moves quickly, and durable-execution and workflow engines (Temporal) may absorb the use cases with LLM-matched wake conditions as an add-on.
Strengths
Limitations
Overall
This is a thoughtful, well-written systems and position paper with modest empirical support. Its value lies in framing and design vocabulary rather than demonstrated performance. Citations are plausible from the agent-infrastructure community, but it is unlikely to become a foundational reference without a release or production evaluation.
Generated Nov 4, 2026
Comparison History (0)
Timely, well-argued systems design with candid limitations and reusable concepts, but evidence is small synthetic pilots and the headline cross-vendor claim is untested