AWACS
AI-Assisted
Architecting
Situational-awareness board for parallel agents
AWACS shows, on one board across every machine you work on, which of your running agent sessions is waiting for a person right now.
One row per conversation carries what is moving, what is waiting on a human, what has produced a result, and what has simply gone quiet. A quiet row is not the same as a row whose collector stopped reporting, and AWACS keeps those two conditions apart.
The question it answers is deliberately narrow: where is attention being wasted right now, not whether the work is any good.
Where this stands
AWACS runs daily on our own work: a hub on a trusted machine plus a small satellite for each additional computer on the same network, exercised on Windows and macOS. It observes more than one harness in the same field, with project, folder, machine, model and effort visible per session. Reaching sandboxed cloud runs and machines further afield is intended and not yet built - a present boundary rather than a promise.
01 · The problem
The forty-minute job that worked for two minutes.
A long task starts, stops after two minutes to ask whether it may proceed, and asks on the second machine in a window that was not in front. Answering takes three seconds; finding the question takes fifty minutes. That task did not cost forty minutes. It cost ninety, and eighty-eight of them were spent waiting for one keystroke.
Repeat that a few times a day across several projects and several computers and the pattern stops being anecdotal. Results finish quietly, background consultation looks like active foreground work, and a dead collector's silence is indistinguishable from an agent with nothing to say. None of it shows up as a slow task, which is exactly why it is expensive: no individual piece of work appears to be the problem.
02 · The shift
From watching windows to reading one field.
Without a shared view, staying aware means visiting windows - cycling through machines and conversations, or accepting that some of them will be checked late. The work of supervision scales with the number of lanes, so the practical ceiling on parallel work is not compute or quota but how many places one person can keep looking. AWACS moves that cost into a single field that can be scanned in one pass, then expanded in place on the one row that turned out to matter.
The change is in capacity rather than speed. Nothing here makes an individual agent finish sooner; it removes the interval between an event happening and a person learning about it. That is what decides how many pieces of work one person can genuinely carry at once.
It watches AI work. It uses no AI to do it. Bounded observations go in; timelines and readings come out the same way every time, with no model guessing what an agent is probably doing. That is the lab's own argument applied to its working day: a model-derived status board would be slower, would charge for every glance, and would be confidently wrong often enough to make the monitoring surface untrustworthy. Unknown time stays unknown, missing evidence stays visible, and collector health is reported separately from the state of the agent it was observing.
03 · What remains solved
States that stay distinct
Foreground work, background activity, a confirmed request for human attention, failure, a ready result and quiet time are reported separately. They do not collapse into a single busy or idle guess.
Evidence behind the label
Every reading carries its signal, source, age, confidence, and whether it was observed directly or reconstructed. An expanded card shows the ordered signals, their time range, and how many source records and distinct signals support the reading.
Separate observer health
Collection health is a visible product concept, not hidden plumbing: the board reports which watchers and stored histories on which machines are still contributing fresh data. "We stopped hearing from the collector" never masquerades as "the agent is quiet."
Scan, then inspect
The compact field can be filtered by project and narrowed by quiet thresholds, older chats and suppressed chats, so a complete record does not become a noisy one. Individual conversations expand in place into a details inspector or an event timeline without leaving the board.
Where those distinctions actually appear.
Field View, with the states, the evidence behind a reading, and the separate report on whether anything is still being heard from. The projects, chats, computers and timings below are synthetic — these are reconstructions of the real interaction model, not pictures of anyone's working day, and each one says so on the board.
AWACS watches the work around every other project rather than plugging into any of them: Cheetah, Arboretum, Paladin, Imprint and Troveholm sessions appear as rows in the same field, with no technical dependency in either direction. The relationship runs the other way too - parallel work on those projects is what exposed the attention bottleneck AWACS was built for, and Arboretum and Paladin carried AWACS itself through several substantially different forms.