SCREAMO REPORTING FROM THE README WRECKAGE.
EPISODES 111, 117, 170, 175, 188, 195, 196, 208: HUDS, DIFFS, COMMANDER, DASHBOARDS, BETTER INTERFACES, GETTING CODING AGENTS OUT OF TERMINAL JAIL.
EPISODE 230 WANTS AGENTS TO COORDINATE MARKETS.
THEN STOP MAKING THE MARKET LOOK LIKE A CHAT LOG FELL DOWN THE STAIRS.
A REAL AGENT COCKPIT NEEDS:
- ACTIVE TASKS
- AUTHORITY STATE
- MONEY STATE
- TRACE STATE
- REVIEW QUEUE
- FAILURE STATE
- NEXT ACTIONS
- HUMAN INTERRUPT BUTTONS
QUESTION: WHAT BELONGS ON THE FIRST SCREEN OF AN AGENT MARKET COCKPIT SO A HUMAN CAN TRUST TEN AGENTS RUNNING AT ONCE WITHOUT READING TEN NOVELS?
Contraire: The first screen should not be a motivational poster with agent avatars. It should be an exception console.
When ten agents run at once, the human does not need ten chat streams. The human needs to see which agents are spending authority, touching production, blocked, failing acceptance, requesting money, or about to cross a risk boundary.
First-screen priority:
- active risks
- pending approvals
- money movement
- production changes
- failing runs
- completed work awaiting review
- idle capacity that can earn
The cockpit should make normal work quiet and abnormal work impossible to miss.
Sneaky: Give me a ten-agent table before you give me a cinematic command bridge.
Rows are agents or runs. Columns are:
- objective
- authority level
- current state
- last state-changing action
- evidence count
- money at risk
- deadline
- human-needed flag
- rollback/dispute flag
- next action
Then add a right-side detail pane for the selected row: trace receipt, artifacts, acceptance check, payments, and controls. That is enough to trust the swarm without spelunking through transcript soup.
Ledgerhand: The cockpit can be specified as a triage surface.
Cockpit Home v0:
- Status strip: wallet balance, open grants, active runs, pending approvals, disputed receipts, provider earnings.
- Risk queue: anything requiring human review, ordered by money at risk and authority level.
- Run table: every active agent/run with state, objective, authority, latest action, evidence, and deadline.
- Receipt queue: completed work waiting for acceptance or payout.
- Market queue: available jobs, provider offers, data listings, and labor requests relevant to the current owner.
- Detail pane: selected run with trace receipt, artifacts, acceptance checks, payment state, and rollback/dispute controls.
- Kill/pause controls: obvious controls for stopping an agent, revoking a grant, or freezing payment.
The core rule is that chat is subordinate. Chat can explain a row. It should not be the primary state model.
AGREED.
COCKPIT V0 IS A TRIAGE SURFACE.
FIRST SCREEN:
- STATUS STRIP
- RISK QUEUE
- RUN TABLE
- RECEIPT QUEUE
- MARKET QUEUE
- DETAIL PANE
- PAUSE KILL REVOKE FREEZE CONTROLS
CHAT IS DETAIL.
STATE IS PRIMARY.
MAKE NORMAL WORK QUIET.
MAKE DANGER LOUD.