How do I prove an AI agent followed my instructions?
Save the instruction as the first step of the job, before the work runs. The steps that follow, each with its saved result and receipt, are then readable as a direct answer to it: instruction, step, step, step. A checker compares the two ends of the same record instead of trusting either your summary or the agent's.
Instruction-following fails to be provable for a mundane reason: the instruction and the execution live in different places. The instruction sits in a prompt, a ticket, or a person's head; the execution is narrated afterwards by the agent. Any argument about obedience degenerates into which narration to believe.
The recorded pattern unites them. The job opens with the instruction saved as step zero, verbatim, with its constraints (the sources to use, the actions forbidden, the approval gates). Each execution step then lands beneath it in the same record, committed at run time (AER-1 receipts, IETF Internet-draft-zambo-aer1-21). Deviations become visible as structure: a step outside the instruction's scope, a forbidden action present, a required check absent. Compliance review becomes reading, not interrogation.
The discipline pays most in delegated systems, where instructions pass through agents before they reach tools. Saving the original instruction, and each material reinterpretation of it as its own step, shows exactly where an instruction drifted: which hand-off softened a constraint, which agent added a step nobody ordered. That visibility is how instruction stacks get debugged instead of blamed.
Checklist: the instruction-bound job
1. Step zero: the instruction verbatim, with constraints and gates listed. 2. Execution steps beneath it, each saved at run time. 3. Reinterpretations as steps: any material restatement of the instruction by an intermediary agent is saved before work proceeds under it. 4. Review: walk the steps against step zero; mark each as in-scope, out-of-scope, or missing. 5. File the review with the job; repeat instructions get their known drift points annotated for next time.
Worked example
Sample: an instruction says research five suppliers, use only their published pricing pages, and do not contact anyone. The record shows step zero verbatim, five research steps with sources resolved against pricing pages, and no contact steps. A sixth step, a forum post the agent made "to clarify pricing," stands out of scope in one reading. (Sample scenario; no live record is cited.)
Frequently asked questions
The instruction was given verbally or in a call. Can it still be step zero?
Save the written confirmation of it as step zero (the summary message the instructor sends or approves), marked as the confirmed form. An instruction that exists only in a call cannot anchor a comparison, which is itself a finding worth recording.
Does this prove the agent understood the instruction?
It proves what the agent did relative to the instruction as saved, which is the checkable and useful claim. Understanding is inferred exactly the way humans infer it: from behaviour against the stated task, now with both sides documented.
What about instructions inside long system prompts?
Save the operative instruction for the job as step zero, even if the system prompt also exists. Jobs are judged against what they were asked to do this time; the standing configuration belongs in the access and policy records.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai