ALGEDI SI Research Labs
Menu

Research program B

Agentic systems and execution reliability

How AI agents perform multi-step work against real systems, and how to verify what they did rather than what they report.

An agent’s final message is a claim. The system it acted on holds the evidence. This program treats every reported outcome as a hypothesis to be checked against files, state transitions and side effects, and studies the conditions under which the two diverge.

Open questions

  1. How often does an agent's report of success match the actual state of the system it acted on?
  2. Which completion invariants can be checked independently of the agent's own account?
  3. Under retries and partial failure, when do agents produce duplicate or inconsistent side effects?
  4. What is lost when long-running work is handed from one context or agent to another?

Research in this program

ALG-002
Publication status: Research question

Agent completion verification

When an agent reports that a task is complete, how often does independent evidence agree? This entry frames the question and the checks that would settle it. It reports no results.

  • Program B · Agentic execution
  • Program D · Evaluation methodology
  • No artifacts yet
Next program: C · Human-AI interaction over time