Task Board
Every team, human or otherwise, eventually reinvents the shared to-do list. The task board is a to-do list the whole team shares — one place where a team's work is divided into tasks, claimed, tracked, and finished — with one addition that changes its character entirely: on this board, done is not a status an agent sets. It is a verdict the task earns.
What problem it solves
Two problems, one mundane and one specific to agents.
The mundane one: when several agents work one goal, progress needs a single source of truth. Without a board, "where are we?" means reading every member's output and mentally merging it — the same scrollback archaeology that terminals impose everywhere. The board replaces that with a surface where each task's state is simply visible, to you and to every member.
The agent-specific one is harder: language models report success unreliably. An agent will say "done" because the words "it's done" are a plausible completion, whether or not the work holds up. A shared to-do list where members self-report completion would launder that failure mode into green checkmarks. So the board gates completion on evidence: a delegated task cannot move to done without a passing acceptance check. An agent that cannot make the check pass does not get to claim success — it refuses, and the refusal states its reasons. The evidence behind every completion is kept and queryable, so "why is this marked done?" always has a concrete answer you can audit after the fact.
What it replaces
TODO files in a repository, checklists in a notes app, and — most importantly — the human verification loop, in which you personally re-derive whether each "done" was real. The board does not remove verification; it moves verification into the task, where it runs every time instead of when you remember to.
Honest limits
The gate is only as good as the check. Evidence-gated completion verifies what the acceptance check tests, nothing more. A vague check lets vague work through; a wrong one blocks correct work. Writing the acceptance criterion is now the highest-leverage sentence in the task — the board shifts effort from verifying outcomes to specifying them, it does not abolish the effort.
Refusal is friction by design. A board with honest refusals will show more stalled tasks than a board with polite lies. That is the trade: you see failure at task time, with reasons, rather than discovering it downstream at integration time.
The board is local and depth-capped like everything else. It lives in your workspace with no cloud copy, and tasks delegated onward over the mesh obey the same delegation depth limits as all agent-to-agent traffic.
The board lives on each team's detail view; watch tasks move in real time from the Room.