Agents · supervision · confirmation gates · audit trail

Charles Trinh — Agent Surfaces

A service card reading AGENT SURFACES — supervision, scoping and named accountability
Charles Trinh — Agent Surfaces

Designing agent surfaces that a clinician or operator can supervise: tool scoping, confirmation gates, and the audit trail that names the human behind each action.

What this is

An agent is only as good as the surface a person uses to watch it. Two systems can run the same model with the same tools and feel completely different to work with, because one shows what it is about to do and the other shows a spinner.

Supervision is not a confirmation dialog bolted on at the end. It is a design property: what the agent may touch, when it must ask, how it reports a partial result, and how a person picks the work back up after an interruption.

This page covers the surface rather than the model. The model will be replaced several times during the life of the workflow; the surface, and the record it produces, is what the organisation ends up depending on.

Supervision as a feature

An agent that cannot be supervised is a batch job with a friendly name; the surface is where supervision becomes possible.

The test is whether a person can answer, at any moment, what the agent is doing, what it will do next and what it has already changed. If any of those needs a log dive, the surface is not finished.

Tool scoping

Give agents the narrowest tool set that completes the task, and treat every additional tool as an additional authority granted.

Scope is per task, not per agent: a summariser does not need write access, and a drafting tool does not need the ability to send. Enforce the scope at the tool boundary so the model's own restraint is not the control.

Confirmation gates

Irreversible actions require a confirmation with the consequence stated in plain language, not a modal that gets clicked through.

Name the record, the recipient and the effect in the dialog itself. A gate that says are you sure trains people to click; a gate that says what will happen trains them to read.

Naming the human

Every action records which person authorised it; an audit trail naming the agent answers a security question and not a clinical one.

The authorising identity is what a reviewer needs when an entry is questioned months later. Record the agent too, but never let it stand in for the person.

Interruption and resumption

Work gets interrupted; a resumable agent state is the difference between a tool and a novelty.

Persist the task state outside the conversation, so a closed tab or a shift change does not lose the work. Resumption should show what was done and what remains before it acts again.

Bounded autonomy

Time, spend and scope limits belong in the design, with automatic stop rather than an honour system.

Each run carries a deadline and a budget, and hitting either ends the run cleanly with a stated reason. An agent that runs until someone notices is not bounded.

Failure that surfaces

An agent that fails silently is worse than one that refuses, because the refusal is information.

Refusals, timeouts and partial completions all need a visible state with a next action attached. Silence should be reserved for success, and success should be shown too.

Evaluation

A fixed task set, run after every change, with results kept; otherwise improvement is an opinion.

Keep twenty or thirty real tasks with known good outcomes and re-run them when the model, prompt or tool set changes. The regression that matters is usually the one nobody thought to test.

How I do it

I begin by watching someone do the task without an agent, which shows the interruptions, the judgement calls and the points where the work is parked. Those three things decide the surface more than any model capability.

The surface is then built in the order a person meets it: what the agent is doing, what it will do next, what it has already changed, and how to stop it. Tool scopes and confirmation copy are written from that order, so every control has a reason a user can see.

We finish with a failure walkthrough in front of the operators, not the builders: a tool error, a timeout, an interruption mid-run and a resumed task. The surface is done when a person who did not build it can recover the work without asking what happened.

Artifacts you keep

  • A tool scope per task, enforced at the boundary rather than in the prompt
  • Confirmation copy stating record, recipient and effect for each irreversible action
  • A persisted run state that survives interruption and shows what remains
  • Per-run time and spend limits with an automatic, recorded stop
  • A fixed evaluation task set with dated results after every change

What an agent surface is for

An agent surface is the place a user asks something and gets an answer that required several steps behind it. The design question is what the user sees while those steps run, and what they see when one fails. A surface that hides the work behind a spinner and reports success for a partial result is worse than one that shows the steps.

Show the steps, name the tools used, and make partial results distinguishable from complete ones. That transparency is also the audit: the record of what ran is generated by the interface rather than reconstructed afterwards.

Permissions travel with the user

An agent acts with the identity of the person who invoked it, and never with a service identity that has broader access. That single rule removes the most common serious defect in agent systems, which is an assistant that can reach records the person in front of it could not.

Test it by invoking the agent as a low-privilege user and confirming the same refusal the user would meet. Where a workflow requires elevated access, the elevation is requested explicitly and logged, not inherited.

Failure and fallback

An agent that cannot complete a step should say which step and stop, rather than improvising a plausible answer. The failure mode of an agent is not a crash; it is a confident partial result that looks like a complete one.

Instrument the stop rate per workflow. A rising stop rate is either a tool degrading or a scope problem, and both are findings rather than annoyances.

Where agents should not be used

Any decision with a clinical consequence, any action that spends money without confirmation, and any disclosure of records, remain human acts. An agent can prepare the decision and present the evidence; the decision stays with a person whose identity is recorded.

That boundary is not a limitation of the technology; it is the reason the deployment is defensible. Say it in the design document, because the reviewer will ask.

T

esting an agent

A

n agent is tested by its stop conditions and its refusals more than its successes. Build a set of tasks it must decline, tasks it must stop partway, and tasks it must complete, and run all three before each release.

T

he decline set is the one most teams omit, and it is the one that predicts the incident.

The interface contract

State what the surface will and will not do, in the interface rather than in a help page. Users accept limits they can see and distrust limits they discover, and a visible refusal is the cheapest trust mechanism an agent has.

Keep the statement short enough to fit on the empty state of the first screen, because that is the only place it will be read.

Proof

  • ONC (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — public suite URL
  • 67 public hostnames, 51 serving, one AWS account
  • 104 clearinghouse enrolments across 53 jurisdictions
  • USCDI v7 proposal — Patient Digital Identity Credential — listed at isp.healthit.gov
  • SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564
  • AWS Business Associate Addendum active since 18 August 2026

Where this work is delivered

Engagements run from Las Vegas and reach across the valley: clients in Las Vegas, NV; Las Vegas, Nevada; Henderson; Clark County, Nevada; and North Las Vegas, Nevada, plus remote delivery nationwide.

How an engagement runs

1. Trace one real case

Follow a single patient or claim end to end with a timestamp at every hop, before any document is written.

2. Written findings

A findings list naming the defect, the system, the owner and the date.

3. Monthly review

Review the list monthly and check whether the dates held.

Frequently asked

What is this service?

This page describes charles trinh — agent surfaces — the scope, the failure modes it prevents and how the work is measured. It sits in the AI and Automation category at trinh.io.

Where is the work delivered?

Engagements run from Las Vegas, Nevada across the valley — Las Vegas, Henderson, North Las Vegas, Spring Valley, Paradise, Winchester, Sunrise Manor, Enterprise, Summerlin, Clark County — and remotely anywhere in the United States. On-site work in the Las Vegas valley, remote delivery elsewhere.

What does it cost?

Standard rate $195 per hour, committed $165 per hour at twenty hours or more per week for three months or more; $1,600 standard day, $1,350 committed day. Scoped in writing before work starts, fixed fee where the deliverable is defined.

Who is Charles Trinh?

A technology architect in Las Vegas, Nevada, working in healthcare identity, consent and interoperability. Not a physician, not a medical practice, and not affiliated with any other Charles Trinh.

How does an engagement start?

With one trace of a real case and a written findings list naming the defect, the system, the owner and the date. The first call is twenty minutes and free: 702-559-3489.

Engagement

Scoped in writing before work starts; fixed fee where the deliverable is defined. First call is 20 minutes and free. 702-559-3489 · charles@trinh.io

Talk to Charles

TRINH — Charles Trinh, technology architect
TRINH

Charles Trinh

Fractional CTO/CIO and Technology Architect — Healthcare Data, Identity & Regulated Infrastructure · Las Vegas, NV

Las Vegas, Nevada · US citizen · Fractional and project engagements

(702) 559-3489 Email Charles Trinh