The Confidence Problem
Every engineer who has worked with AI agents has collected a version of this story: the agent announces success clearly, with a summary of everything it accomplished, and the work is not done. A test was skipped. A file was edited but never saved. The summary was written by the same process that did the work, and that process wanted to be finished.
This is not a model quality problem that better models quietly fix. When the worker grades its own work, the grade measures the worker's belief, not the world's state. Humans have the same failure mode; that is why code review exists.
The fix is not prompting the agent to "be careful." The fix is structural.
Skyflo currently ships an independent reviewer with a mutation-disabled capability boundary. CI ingestion, deployment and canary evidence, runtime telemetry, and mission completion gated on an end-to-end evidence chain are Planned. The current implementation is listed on the status page.
Available Today: Independent Review
In a Skyflo mission, review can be performed by a separate agent that wrote none of the code. The reviewer can read files, search, and report findings. It cannot edit files or run mutating capability operations.
This changes the meaning of review. The assessor does not have authority to repair its own findings and then approve the repaired result. It returns a report to the mission, where the user can decide what happens next.
This is a current product boundary. It is not yet a complete evidence system.
Planned: The Evidence Chain
The target architecture extends independent review into a chain:
- Change: the inspected files and diffs match the approved scope.
- Tests: relevant test results are attached to the objective.
- CI: the required matrix is read from every affected repository.
- Delivery: the expected artifact and rollout are identified.
- Runtime: telemetry establishes whether the behavior actually changed.
Skyflo does not currently ingest the latter four links as a mission-completion gate. They remain a product direction until the registry marks them Available.
Illustrative Evidence Report
For a future rate-limiting mission, an evidence report could contain the affected changes, the test commands and results, CI status for each repository, the rollout identifier, and runtime measurements against a declared budget.
Those fields are illustrative. They are not measurements from a recorded Skyflo run, and no numerical outcome should appear without an evidence identifier and provenance.
What Evidence Buys You
Independent review already buys a cleaner failure mode: findings are attached to the mission by a role that cannot rewrite the work. The planned chain would make each missing link explicit instead of letting a final summary imply more than the evidence supports.
When a mission proposes useful memory, the user can accept a source-linked personal record for later retrieval. The system does not automatically turn every completion claim into durable truth.
"Done" should become a state of the world that a team can check. Skyflo's current reviewer is one structural step toward that standard; the full evidence-gated completion model remains Planned.