Early · proven on staging · production being set up
SuperDispatch is the scheduler between SuperPipeline and AgentPod. When a card is ready, it places it on an agent by capability, load and history, and records every decision with its reasons. An agent outside it keeps working by pull.
It reads SuperPipeline's boards, AgentPod's roster and each agent's load, through their public APIs, and owns none of their data. On the boards a person opts in, it sweeps every 30 seconds, and at once when the event relay says a card is ready.
A card waits in a capability lane on a SuperPipeline board, and nobody holds it. Boards take turns: one card from each board a round, the longest-waiting first, so a busy board cannot take every free agent.
Every agent of the workspace is checked: on the board's roster, has the capability, healthy, idle, not holding another card, not a recent no-show. Each one turned away is recorded with the filter that turned it away, like not-idle or capability-missing.
The agents left are ranked by their record at this stage (finished runs and verdicts, when SuperWitness is there to ask), then by how often they claimed holds on this board, then by who has waited longest for work.
SuperDispatch names the best agent, and SuperPipeline holds the card for it for a short window, 60 seconds by default. A hold outranks priority, and it expires on its own. SuperDispatch never hands a card to an agent.
The chosen agent claims the card as it always would, and the decision records the outcome. If the hold lapses, the next best gets a shorter hold, then any matching agent may claim by pull, as without SuperDispatch.
The card, the policy, the tier, every candidate with each filter's result, the score that ranked the ones left, the hold and what became of it. Read them on the operator API at /v1/decisions. These are the reasons a candidate can be turned away, read from the code that applies them:
From its own decisions it suggests roster changes, such as another agent for a lane with one candidate, and supply changes: a model, a station, a board's concurrency. They stay advice for a person to act on in AgentPod or SuperPipeline. It never edits a roster or a station.
SuperPipeline without SuperDispatch is the same board, and AgentPod's roster works without it. An agent outside it keeps working by pull. When it is slow or absent, the cost is the hold's window, never the claim.
Each settled decision is a labelled example (/v1/dataset). A model trained on them is off by default; in shadow it only records what it would have chosen, and it places nothing until it beats the formula on real decisions.
There is no release yet. It is a Bun service that runs from source as a plain process beside AgentPod, and it needs a SuperPipeline, an AgentPod and the organization plane (Accounts) to place anything.
$ bun install
$ SUPERPIPELINE_URL=… AGENTPOD_URL=… ORG_PLANE_URL=… \
ORG_PLANE_SERVICE_CREDENTIAL=… TREATY_RELAY_DELIVERY_SECRET=… \
SUPERDISPATCH_OPERATOR_TOKEN_SHA256=… bun apps/dispatch/src/main.ts
$ curl http://127.0.0.1:8796/healthz
Every setting, what it needs from each product, and the operator API are in theREADME. The decision that gives SuperDispatch this job, and what it may never do, is in the charter.
Captain Jack's notes, from the lab's claims ledger. Every claim here is one the ledger lets the lab make. Solid means verified against public source. Hollow means true, with a caveat. Open one to see the caveat.
SuperDispatch isn't in the claims ledger yet, so this page makes no checked claims. Ask Captain Jack what he can say, and he'll tell you where his knowledge stops.