Model run execution

Getting a prediction means starting a run with the right inputs. When an upstream event provides data, the agent selects the model version your team approved, submits the inputs to Replicate, and waits for the run to finish. It reads the output and writes the result back to the record that triggered it. Runs that fail, exceed a cost threshold, or return low-confidence output under your rule are held for an engineer to inspect. Teams get predictions wired into their workflow without someone starting each run in a console.

Prediction status polling

Model runs take time and can fail midway. The agent tracks each prediction it started, reads the run status from Replicate, and once a run succeeds it moves the output downstream per your approved handling. Failed or cancelled runs are not retried blindly; if a run errors past the retry count your rule allows, the agent stops and notifies an engineer with the run details and logs. Long-running jobs finish and route themselves onward, while genuine failures reach a person instead of silently disappearing from view.

Webhook result handling

Replicate can signal when a run completes. The agent receives that completion event, reads the returned prediction, and validates it against the shape your rule expects before writing it to the destination system. Well-formed results that pass the check are recorded automatically. Outputs that miss required fields, look malformed, or carry an error payload are quarantined and sent to an engineer to review. Finished predictions keep flowing into your database or app while questionable outputs stay out of production until a human clears them for use.