Agents for Humans: the interrupt is the product
An agent that always says yes has not been tested. Why my approval parser is strict, and what happens to the cases when nobody decides.
Last post I argued that the deadline math in Instanter never touches the model, because counting seven days under O.C.G.A. 44-7-51(b) has no judgment call in it.
This one is about the other end. The moment the agent stops.
Instanter does background triage for eviction-defense clinics. It reads an intake queue of Fulton County dispossessory cases, computes every answer deadline from the statute, ranks by proximity to a default judgment, and then does the thing most agent demos skip. It refuses to act, and asks a licensed attorney.
I have come around to thinking that refusal is the product and everything upstream of it is plumbing.
An agent that always says yes has not been tested
The failure I worry about is not the agent being wrong. It is the agent being agreeable.
A human-in-the-loop step that treats any reply as consent is theatre. If a tired attorney in a courthouse hallway types "ok sure" or hits return on an empty box, and the run reads that as authorization to commit a legal filing decision, then the loop exists in your architecture diagram and nowhere else.
So the approval parser is strict to the point of rudeness. Only an exact approval approves: approve, approved, approve all. A deferral has to be explicit and carry a reason. Everything else comes back invalid and the run does not proceed on it.
You can try this yourself right now, no account, no key. The console has a third button sitting next to Approve and Defer that sends a deliberate typo.
That is not a bug I forgot to fix. It is the demo.
Send the typo and the run reports FAILURE. Nothing crashed. Nobody decided.
What happens to the cases when nobody decides
Here is where it gets more interesting than a validation check.
If an agent refuses an ambiguous answer and then drops the work, it has traded one failure for a worse one. The cases it was holding are the ones closest to a default judgment. Silence is not safe here. A tenant is not one day further from a writ of possession because a workflow got confused.
So the refusal has a floor under it. When no valid decision arrives, a deterministic path commits the escalations for later human review, marks them as committed under floor authority rather than under an approval, and the run still reports failure.
The work survives. The authority is not forged.
Those two being separable is the whole design. In the run context an approval and a floor commit are different states with different provenance, and the runner is the only thing allowed to authorize the floor. A tool cannot quietly promote itself.
Binding the approval to what was actually on screen
The subtlest defect I hit in this area came out of an adversarial review round, and I would not have thought of it on my own.
An attorney is shown two cases with rationales. They read them and approve. Between the moment those cases rendered and the moment the approval lands, the underlying queue can change. A rationale gets rewritten. A case gets added. A ranking shifts. The approval then applies to something the human never saw.
The hook computes a digest of exactly what was presented and verifies it when the approval arrives. If the state moved, the approval is refused with STATE CHANGED DURING APPROVAL rather than applied to different content. There is a companion guard for the empty case, NOTHING TO APPROVE, which fires when no interrupt-now cases exist, so a run cannot record an approval of nothing.
Getting that right took a correction I am glad I caught. My first version computed the digest at approval time instead of verifying one recorded at presentation time. That version could never fail. It was comparing the state to itself.
A guard that cannot fail is not a guard. Writing the failing case first is what exposed it.
Rationing attention instead of scoring it
The last piece is capacity. A clinic does not have infinite attorney minutes, so the agent is told how many interrupts it may spend and it spends them on the cases closest to default.
I deliberately did not build a score. One blended number would let a merely urgent case outrank a genuinely overdue one because it did well on some other axis, and it would be impossible to explain to the person you just interrupted.
The ladder is ordered and each level says what it is. Interrupt now. Surface today. Monitor. Hold. A case that is one day past its effective deadline sits at the top because it is past its deadline, not because it accumulated points.
The AWS part
The agent runs as a Strands Agents Graph on Amazon Bedrock AgentCore Runtime. The interrupt is a BeforeToolCallEvent hook that calls event.interrupt(), and the run's state persists so the answer can arrive from a completely different process.
I proved that cross-process resume in a spike before committing to the architecture. A human-in-the-loop pause that cannot survive the compute going away is not a pause. It is a timeout.
That persistence is what makes the phone app possible. The escalation fires while the one person allowed to decide is standing in a hallway at the courthouse, and the decision comes back from there.
Next post: how a browser with no credentials reaches an agent whose only invocation path is IAM signed, and why the sweep now runs at 7am whether or not anybody opens the page.
Instanter drafts only and never files, computes deadlines and never advises, and keeps a licensed attorney in the loop for every case it surfaces. All demo data is synthetic. The statutes are real and public. Built solo for the Agents for Humans Hackathon.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article