Point 1 is the whole game, and I'd put it harder: an approval is a signature, and a signature over a summary is a signature over nothing. The payload is the act — this customer, this amount, this destination — and anything less specific is the approver trusting the agent's compression of its own request, which is the one step that must never be trusted. My house rule, learned from running agents that spend and send on a human's behalf: the approval binds to the exact content shown, and if the content changes after approval — one digit, one recipient — the approval is void and the new version goes back in front of the human before anything moves. Teams resist this because it adds a round trip. The alternative is discovering, during the postmortem, that your audit trail proves someone approved a sentence about the action rather than the action.
"A signature over a summary is a signature over nothing" is a sharper way to put it than I managed. Voiding on any change, down to one digit or one recipient, is the rule that makes the record mean something. The round trip is the real cost; do you find people accept it more easily when the re-approval shows only what changed?
Yes — in my experience the diff is what makes re-approval acceptable, because it shrinks the re-check to what actually moved. A full-document re-approval trains the opposite habit: nobody re-reads forty unchanged pages, so the signature quietly becomes a rubber stamp over everything, including the change. What I send back is the changed lines plus one explicit line stating what did not change — scope, recipients, totals — so the signature still covers the whole record, but attention lands on the delta. The failure mode I watch for: a change described by its author as 'formatting only' that smuggles substance. So the diff has to be machine-generated from the two versions, never the author's own summary of what they touched.
A machine generated diff plus one explicit line on what did not change is a strong pattern: attention lands on the change, and the signature still covers the whole record. The "formatting only" change that hides substance is exactly why the diff can't come from the author. Do you show whitespace and reordering in that diff, or filter them out before the approver sees it?
Binding the approval to a canonical payload is necessary, but I’d also capture the policy evaluation snapshot that produced the approval requirement. A replay then proves not only what was approved, but why that action was classified as needing this approver and this threshold. That is especially useful when the policy evolves between approval and audit.
Capturing the policy evaluation that produced the requirement is a strong addition: it answers why this action needed this approver at this threshold, not only what was approved. That matters most exactly when the policy changes between approval and audit. Would you store the policy version and the inputs it was evaluated on, or the full rule text each time?
For retries, I’d keep the approval ID separate from an idempotency key for the action, then log both with the execution result so a safe retry can’t look like a second approval. Have you settled on a retention period for those records?
Keeping the approval ID and the idempotency key separate is a good call, so a safe retry never reads like a second yes. On retention, the answer that came out of the replies was to keep the proof (who approved, when, the hash, the outcome) for as long as the action can still be disputed, and to drop the full message bodies sooner. Would you keep a different window for actions that can be undone?
The point about re-checking the payload right before it runs is the one I'd keep for customer replies: an approval from an hour ago doesn't prove the message in front of me now is still the right one. For anything going to a customer, my non-negotiable is reading the exact text once more before it sends — if a name, amount, or promise changed since approval, it goes back to a human.
Reading the exact text once more right before it sends is the rule several people here landed on: the approval covers the message as it stands, not as it stood an hour ago. Name, amount and promise is a good short list of what sends it back. Do you treat a change in tone the same way, or only those three?
The gap audit is probably more useful than it sounds. When building automated tools, logging what was blocked matters just as much as logging what went through. Otherwise you're only seeing half the story.
A technologist working in automation. Here I write up what I actually learn building tools, engineering AI agents, and keeping systems running, along with the workflows and habits that stick.
Enforcing limits at the credential boundary is the only control that holds up when the agent harness itself has a bug. An approval button in a chat app only proves someone clicked yes at a point in time, whereas a scoped token or an outbound proxy physically rejects anything outside the narrow rule even if the model loops or gets prompt-injected.
On question 3, the cleanest split is keeping the raw credential out of the agent process altogether. Give the agent a proxy endpoint or a narrow OAuth scope with a hard dollar cap, an allowlist of destination domains or recipient addresses, and an idempotency key per approved payload. When the signing proxy runs outside the agent's container and logs every request and response independently of the model's session transcript, you also get the answer to mickyarun's gap audit. Any outbound action without a matching approval ID fails at the proxy with an HTTP 403 and leaves a rejection line in the proxy log, which removes any reliance on the agent reporting what it skipped.
Keeping the raw credential out of the agent process is the strongest answer to question 3 so far, and the proxy log doubling as the gap audit is a neat connection: any outbound call without a matching approval id is the finding, recorded by something the model can't edit. Have you run a signing proxy like this in practice, or is it the design you'd reach for first?
We're a place where coders share, stay up-to-date and grow their careers.
Point 1 is the whole game, and I'd put it harder: an approval is a signature, and a signature over a summary is a signature over nothing. The payload is the act — this customer, this amount, this destination — and anything less specific is the approver trusting the agent's compression of its own request, which is the one step that must never be trusted. My house rule, learned from running agents that spend and send on a human's behalf: the approval binds to the exact content shown, and if the content changes after approval — one digit, one recipient — the approval is void and the new version goes back in front of the human before anything moves. Teams resist this because it adds a round trip. The alternative is discovering, during the postmortem, that your audit trail proves someone approved a sentence about the action rather than the action.
"A signature over a summary is a signature over nothing" is a sharper way to put it than I managed. Voiding on any change, down to one digit or one recipient, is the rule that makes the record mean something. The round trip is the real cost; do you find people accept it more easily when the re-approval shows only what changed?
Yes — in my experience the diff is what makes re-approval acceptable, because it shrinks the re-check to what actually moved. A full-document re-approval trains the opposite habit: nobody re-reads forty unchanged pages, so the signature quietly becomes a rubber stamp over everything, including the change. What I send back is the changed lines plus one explicit line stating what did not change — scope, recipients, totals — so the signature still covers the whole record, but attention lands on the delta. The failure mode I watch for: a change described by its author as 'formatting only' that smuggles substance. So the diff has to be machine-generated from the two versions, never the author's own summary of what they touched.
A machine generated diff plus one explicit line on what did not change is a strong pattern: attention lands on the change, and the signature still covers the whole record. The "formatting only" change that hides substance is exactly why the diff can't come from the author. Do you show whitespace and reordering in that diff, or filter them out before the approver sees it?
Binding the approval to a canonical payload is necessary, but I’d also capture the policy evaluation snapshot that produced the approval requirement. A replay then proves not only what was approved, but why that action was classified as needing this approver and this threshold. That is especially useful when the policy evolves between approval and audit.
Capturing the policy evaluation that produced the requirement is a strong addition: it answers why this action needed this approver at this threshold, not only what was approved. That matters most exactly when the policy changes between approval and audit. Would you store the policy version and the inputs it was evaluated on, or the full rule text each time?
For retries, I’d keep the approval ID separate from an idempotency key for the action, then log both with the execution result so a safe retry can’t look like a second approval. Have you settled on a retention period for those records?
Keeping the approval ID and the idempotency key separate is a good call, so a safe retry never reads like a second yes. On retention, the answer that came out of the replies was to keep the proof (who approved, when, the hash, the outcome) for as long as the action can still be disputed, and to drop the full message bodies sooner. Would you keep a different window for actions that can be undone?
The point about re-checking the payload right before it runs is the one I'd keep for customer replies: an approval from an hour ago doesn't prove the message in front of me now is still the right one. For anything going to a customer, my non-negotiable is reading the exact text once more before it sends — if a name, amount, or promise changed since approval, it goes back to a human.
Reading the exact text once more right before it sends is the rule several people here landed on: the approval covers the message as it stands, not as it stood an hour ago. Name, amount and promise is a good short list of what sends it back. Do you treat a change in tone the same way, or only those three?
The gap audit is probably more useful than it sounds. When building automated tools, logging what was blocked matters just as much as logging what went through. Otherwise you're only seeing half the story.
Exactly. The blocked list is often where the useful signal is: it shows what the agent tried, and whether the rules are too tight or doing their job.
Enforcing limits at the credential boundary is the only control that holds up when the agent harness itself has a bug. An approval button in a chat app only proves someone clicked yes at a point in time, whereas a scoped token or an outbound proxy physically rejects anything outside the narrow rule even if the model loops or gets prompt-injected.
On question 3, the cleanest split is keeping the raw credential out of the agent process altogether. Give the agent a proxy endpoint or a narrow OAuth scope with a hard dollar cap, an allowlist of destination domains or recipient addresses, and an idempotency key per approved payload. When the signing proxy runs outside the agent's container and logs every request and response independently of the model's session transcript, you also get the answer to mickyarun's gap audit. Any outbound action without a matching approval ID fails at the proxy with an HTTP 403 and leaves a rejection line in the proxy log, which removes any reliance on the agent reporting what it skipped.
Keeping the raw credential out of the agent process is the strongest answer to question 3 so far, and the proxy log doubling as the gap audit is a neat connection: any outbound call without a matching approval id is the finding, recorded by something the model can't edit. Have you run a signing proxy like this in practice, or is it the design you'd reach for first?