1. Use AgentFoundry for execution, not vague research
AgentFoundry is for governed autonomous engineering work. If the job is still mostly research, framing, or requirements discovery, start in Lexopedia AI first and move into AgentFoundry once the objective is clear enough for agents to execute.
A good AgentFoundry task says what should change and how the change will be judged. If the only instruction is `make this better`, the run will be harder to review. If the instruction names the affected files, success checks, human approval point, and rollback expectation, the result becomes much easier to trust.
2. Prepare clean task boundaries
A strong AgentFoundry request defines the objective, repository or environment, constraints, approval points, and what counts as done. The clearer the scope, the easier it is to review what engineering agents produced.
Before launching a run, decide whether you want diagnosis, a patch, a test, a written report, or a full implementation. Those are different jobs. Mixing them without priority makes the output less useful and makes review slower.
- State the objective in one sentence.
- Name the repository, branch, or code area involved.
- List constraints: policy, security, budget, deadlines, or files that must not change.
- Define success: tests, report, patch, review note, or explanation.
3. Review the evidence, not just the result
AgentFoundry should make the work reviewable. Do not look only at the final patch or final claim. Review what changed, what passed, what failed, what still needs approval, and what the run cost or consumed.
- Check repository state and scope first.
- Check tests, checks, and review points.
- Check human approval points before accepting or merging anything important.
- Escalate when a run looks outside scope or under-evidenced.
4. Keep humans in the approval loop
AgentFoundry is designed for governed autonomous delivery. That means human judgment still matters at the approval point. Use the system to let agents perform bounded work and improve traceability, not to skip review for risky changes.
5. Ask support when the run model is unclear
Contact support if you are unsure whether a job belongs in AgentFoundry, if the execution state looks inconsistent, if review notes are missing, or if access, billing, or pilot questions block adoption.