Blast reviewer
Walks the live model from the changed node and writes the confidence-ranked comment a senior reviewer would have written. The PR knows what it might break before merge.
Each does one operational job end to end. Each cites its evidence. None of them invent a culprit to look smart. They are the same model you ask questions of, narrowed and automated to one job each.
Each scene below is the agent doing its job end to end. The trigger, the recon, the verdict. The same model underneath all five.
Walks the live model from the changed node and writes the confidence-ranked comment a senior reviewer would have written. The PR knows what it might break before merge.
Page lands and the same brain runs the recon: window, traces, recent deploys, the most likely cause with cited evidence. The on-call reads a verdict, not a smoking pile of logs.
Catches the breaking schema change the moment it lands, walks every consumer from the model, blocks the merge if anyone is on incompatible shape, proposes the shim.
Sweeps runtime usage against the graph. Names the modules paying for capacity nobody calls, and proposes a new shape the model can verify is safe to apply.
Watches declared shape against observed reality. Flags the moment k8s drifts from the manifest, names the divergence, routes to the team the model says owns the node.
Every agent bounds its question before answering. A regression has a window. A change has a radius. A breach has a perimeter. No agent ever queries the whole graph because it does not have to.
Every tool call carries the extractor that produced its result. Every confidence number is real, from the merge. The verdict can be traced edge by edge back to a line of code or a span.
When the evidence is not there, the agent says undetermined. Not maybe. Not probably. The refusal is logged the same way an action would be. Auditable, never hidden.
Agents graduate as the model earns trust. Each percentage point of verified truth raises the ceiling on how much you can let an agent act without a human checking.
Agents answer what is happening with cited evidence and stop there. The human acts.
Agents answer and propose the discrete action with a confidence. The human approves, the action runs.
For narrow, high-confidence cases the agent acts on its own. The human reads the receipt.
Get the v0.1 launch notification and a 30-day head start before public release.