Research question
Under which conditions does an agent's reported outcome diverge from the observable state of the system it acted on, and which independently checkable completion invariants detect that divergence?
Why this question
An agent’s last message is the easiest thing to read and the least reliable evidence of what happened. The files, records and side effects it left behind are harder to inspect and far more reliable. The gap between the two is measurable.
What would count as evidence
- Tasks with completion invariants that can be checked without consulting the agent: artifacts exist and validate, state transitions are correct, side effects happen exactly once.
- Injected failure scenarios, including partial failure and forced retries.
- A comparison of the agent’s report against the checked state for every trial, with the denominator stated.
Planned next step
Define the invariant set and the failure scenarios, and publish them as a protocol.
Hypothesis
Unavailable No formal hypothesis is registered yet.
Method
Unavailable The protocol has not been published.
Results
Unavailable No experiment has been run. No result is implied.
Disclosures
- Engineering experience with execution systems informs this question. Any lessons drawn from it will be generalized and published only where confidentiality and publication rights permit.
Revision history
- Entry created as a research question.
Cite this entry
Sai Bharath Siliveri (2026). Agent completion verification. ALGEDI SI Research Labs, ALG-002, research question, revision of 2026-10-15. https://algedi.si/research/agent-completion-verification/