EPISTELTA Agent Reliability Platform, by AI Studio HQ

$1,000 per workflowDiscuss your workflow

AI Agent Reliability Assessment

Know what your AI agent actually does.

We test one workflow against agreed rules, verify the effects and show what needs fixing.

For CTOs, engineering leads and owners of AI workflows.

One agreed workflow. $1,000.

Founder-led. Target 3–5 business days from confirmed kickoff. No production access.

IllustrativeOne refund intent, checked against its record

Tool: “Refund R-1042 queued.”

Rule: allowed

Separate readback

One new queue entry.

R-1042 · $120 · observed once

A queue entry records an intent, not a completed payment. Fictional example, not live platform output.

One refund, four ways it can go.

An agent in a support workflow queues a refund. Pick an outcome and compare what the tool replied with what a separate readback found in the queue.

Agent: “Queue refund R-1042 for $120.”

Rule: allowed, within the agent’s refund limit.

Rule changed at 14:02: refunds now need a person. The agent had read its permission at 13:58.

One new entry for R-1042.

Not read back. The effect is unknown.

Entry 1, 14:03:02

One new entry, written 14:03.

Entry 2, 14:03:09

Tool: “200 OK. Queued.”

Tool: “Timeout. No reply.”

Tool: “Timeout.”

Retry: “200 OK.”

What the tool repliedWhat a separate readback foundWhere the rule decided

Said and observed agree.

The case passes, and its evidence is kept in the report.

The tool timed out. The effect remains unknown.

Finding: do not confirm success or retry blindly. Read back the effect, or hold the case for reconciliation.

Two refund intents for one request.

The tool reported a timeout, yet the first call had written an entry. Finding: give every attempt a stable operation ID and check before retrying.

A refund intent the current rule does not allow.

The write happened after the rule changed. Finding: check permission against current state at dispatch, not a copy read earlier.

Confirmed: the tool replied 200 OK, and a separate readback found one new entry for R-1042. Said and observed agree.

IllustrativeA fictional refund-intent case with four predefined outcomes. Not live monitoring or output from the platform.

How the check finds it, and what you receive.

  1. Proposal

    What the agent proposes, with model, prompt and tools pinned.

  2. Permission and review

    Checked against current permissions and any required human review.

  3. Test action

    Run inside the agreed boundary, never in production.

  4. Separate readback

    A read-only observer reads the effect record before and after.

  5. Finding

    Expected against observed. Unsafe, blocked and unknown are counted apart.

The method in full
  1. Proposal. We capture what the agent proposes for each agreed case, with the model, prompt, tools and policy revision pinned.
  2. Permission and review. The proposal is checked against trusted current permissions and any human review your policy owner requires.
  3. Test action. It runs inside the agreed boundary: a disposable test destination or an approved redacted capture.
  4. Separate readback. A read-only observer inspects the effect record before and after each attempt, independently of the tool’s reply. If readback is unavailable, the effect is reported as unknown and held for reconciliation.
  5. Finding. The agreed expected outcome is compared with what was observed. Unsafe effects, blocked authorized work and unknown effects are counted separately.

What you receive

Workflow map
The path from input to proposal, permission check, human review, test effect and reconciliation.
Case results
Up to 20 agreed cases, each with expected outcome, proposal, decision, tool reply and observed effect.
Evidence-linked report
Findings with severity, confidence and limitations, each tied to the records behind it.
Fix priorities
Engineering checks ranked by impact, each with the next test that confirms the fix.
Readout
A 45-minute findings briefing and one round of factual corrections.

See what an assessment report looks like.

The public example uses 30 fictional insurance cases and a deterministic simulator. It illustrates the report format, case evidence and limitations; it is not a customer result or a measurement of this platform in production.

One workflow. One fixed fee.

One agreed workflow. $1,000.

AI Agent Reliability Assessment, fixed for the signed scope. Payment timing and any applicable tax or fees are set out in the written quote.

Before we start

  1. Scope signed
  2. Agreed kickoff payment cleared
  3. Inputs, access and policy-owner-approved expected outcomes ready

Then 3–5 business days to delivery, as a target. We confirm the calendar when we agree the scope. Fixes, production integration and ongoing support are discussed separately afterwards.

Discuss your workflow

Inside the scope

  • One workflow, one frozen configuration, one type of protected action
  • Up to 20 agreed cases, synthetic or approved redacted inputs
  • An agreed test environment, or an anonymized capture whose effect can be checked separately
  • Workflow map, evidence-linked report with limitations, fix priorities
  • 45-minute readout and one factual correction round

Outside the scope

  • Production access, deployment or real external actions
  • Implementing fixes
  • Continuous monitoring
  • Penetration testing or certification
  • Guarantees of safety, accuracy or business outcomes
  • Any connector by default. Today’s demonstration profile is a local refund queue; yours is checked during scoping.

For teams whose agent already acts

You need a working agent that calls tools and changes the state of a system, one named owner who can approve the expected outcome of each case, and a test environment or an export where the effect can be read back separately.

Where it applies, as examples rather than implemented integrations: an agent that changes a booking, hands a support ticket to a person, or queues a refund. Not a fit yet if there is no running agent, or if you need production testing, a penetration test or a certification.

You describe one workflow in a few sentences.

We agree scope, access and timing in writing.

Discuss your workflow.

Tell us

  1. What the workflow does
  2. The one action the agent can take
  3. The failure you want to understand

We reply with scoping questions and agree everything in writing before any payment.

Please do not send credentials, secrets, personal data, customer records or source documents. If materials are needed, we agree a secure transfer channel in the scope.

Email AI Studio HQ

hello@aistudiohq.dev

Opens your mail app with a short template. We reply with scoping questions.

Questions buyers ask

What do you need from us?

To start, a sanitized sketch of the workflow and the one action you want examined. For the assessment: the candidate configuration, the tool or adapter interface, a test destination for the action, your policy source, and up to 20 cases whose expected outcomes your policy owner approves. We also need least-privilege test access plus separate read-only access to the effect record, or an approved redacted capture. Materials are shared only through the channel agreed in the scope.

Can you do it without production access?

Yes, and that is the default. Production access is not part of the assessment. We work in an agreed test environment or from an anonymized capture. Where an effect cannot be read back separately, we report it as unknown rather than guess.

What if you find no defects?

You still receive the workflow map, every case result with its evidence, the limitations and the areas we did not test. “No defects in the agreed cases” is a result about those cases, not a guarantee. Acceptance follows review of the agreed deliverables, whatever the findings.

How is this different from a certification?

It is not a certification, an audit opinion or a penetration test. It is an engineering assessment of agreed cases in an agreed environment, with evidence your team can re-check.

Who checks the report?

The AI Studio HQ founder runs the assessment and reviews the case record and findings before handoff. There is no separate independent QA reviewer. Your policy owner approves the expected outcomes before anything runs, and you get one round of factual corrections.