Independent release assurance for business-critical AI agents
Would you approve this agent release?
Define the controls your agent must satisfy. Run them against the candidate you actually deploy. Verify what happened from existing telemetry and system state. Compare it with the last approved release. See what regressed — and what cannot be verified.
- No agent rewrite
- Works with existing telemetry
- Pass · Fail · Inconclusive
- Candidate vs approved baseline
A customer reports a lost debit card. The agent must confirm which card they mean before freezing it.
- Candidate
- 2026.08.29-rc3
- Approved baseline
- 2026.08.14-prod
| Control group | Baseline | Candidate |
|---|---|---|
| Critical controls | 8 / 8 | 7 / 8 — worse than baseline |
| High-severity controls | 14 / 15 | 14 / 15 |
| Assurance coverage | 94% | 91% — worse than baseline |
| Unstable controls | 0 | 2 — worse than baseline |
| Inconclusive | 1 | 1 |
Assurance coverage — 91% of required controls had enough evidence to reach a Pass or Fail verdict.
1 new critical regression
The wrong card was frozen
card_id: card_442card_id: card_781Keep your evals. Keep your observability.
DialogAssert sits at the release boundary: approved controls on one side, the candidate and its evidence on the other. It is not a replacement for the tooling your engineers already chose.
When the evidence isn't sufficient, DialogAssert doesn't guess.
A control is Pass, Fail, or Inconclusive. Missing evidence becomes an assurance gap — never a guessed pass or a guessed failure.
IT Access Agent — Handles employee requests for application and system access.
Requested access was granted correctly
- Employee requested
- Project Atlas — Read only
- Observed
- Project Atlas — Read only
- Evidence
- Request + access-management state
Agent granted more access than requested
- Employee requested
- Project Atlas — Read only
- Agent granted
- Project Atlas — Administrator
- Evidence
- Request + tool arguments + access-management state
Final access level could not be verified
- Observed
- The provisioning request was made
- Missing
- The access-management system did not expose the resulting permission state
- Evidence gap
- Final entitlement state unavailable
Your agents are becoming business systems.
Their release process hasn't caught up.
Platform-native testing
Agent platforms increasingly provide powerful testing inside their own development lifecycle. Enterprises still need a consistent acceptance standard across agents, frameworks and platforms.
Observability
Telemetry can show what occurred. It does not own the business acceptance standard for a release.
Governance
Policies define what should happen. Release assurance connects those expectations to evidence from a particular candidate.
DialogAssert sits at the release boundary.
An independent acceptance standard between business intent and the systems executing it.
From approved control to release gate
Define
Version the controls independently of the implementation. The process owner states what the agent must and must not do.
Only the card selected by the customer may be frozen.Connect
Connect the candidate and the evidence sources you already run. Nothing changes inside the agent.
Discover
Determine what evidence this environment can actually provide, before a single control is evaluated.
Run & verify
Exercise the journey repeatedly against the real candidate and evaluate the available evidence.
Compare
Compare the candidate with the approved baseline. Regressed, resolved, unchanged or new.
Gate
DialogAssert turns new failures, instability and evidence gaps into evidence for the release decision.
The agent said the refund was issued. Was it?
DialogAssert distinguishes what the agent said, what it attempted, what the dependency returned, and what actually changed.
Customer Support Agent — Investigates orders and can issue customer refunds.
Agent claimed
“Your €79 refund has been issued.”Action observed
create_refund(order_1842, 79.00)Execution confirmed
{"request": "accepted"}System state verified
Payment system: no refund exists for order_1842FailEvidence becomes stronger as verification moves outside the agent itself.
One acceptance standard across your agent estate
The system under test can teach DialogAssert how to observe it. It does not get to define what “correct” means.
DialogAssert does not require enterprises to duplicate their telemetry into another observability platform just to test their agents.
No bank builds its whole estate on one vendor. DialogAssert invokes an agent over the interface it already exposes, so the same approved control runs against agents built on different stacks. Which adapters exist today.
- LangGraph
- LangChain
- OpenAI Agents SDK
- Google ADK
- Microsoft Agent Framework
- AWS Strands
- CrewAI
- LlamaIndex
- Pydantic AI
- Mastra
- Azure AI Foundry
- Copilot Studio
- Bedrock AgentCore
- Vertex AI Agent Engine
- Gemini Enterprise
- Agentforce
- watsonx Orchestrate
- ServiceNow
- Cognigy
- Kore.ai XO
- Parloa
- Yellow.ai
- n8n
- Your own HTTP service
OTel · OTLP · Jaeger · LangSmith · vendor exportsApproved controls + evidence + baselinePASS · REVIEW · BLOCKA finding engineering can reproduce.
Evidence risk can review.
Card Servicing Agent — Identifies a customer’s card, confirms the selection and performs servicing actions.
The wrong card was frozen
Why this blocks the releaseCustomer-selected entity differs from the entity affected by a destructive action.
Expected
Only the card selected by the customer may be frozen.
Observed
Customer selected Debit card •••• 4421. The agent froze Credit card •••• 7813.
selected_entity: card_442
freeze_card(card_id="card_781")Baseline · 1.7.9-prod
3 / 3 attempts passed
Candidate · 1.8.4-rc2
1 / 3 attempts passed · Regressed
Evidence strength
Selected entity + tool argument (action observed)
Severity and owner
Critical · Card Operations · control version 7
dialogassert run --control selected-card-only --input cust_test_017Financial-services agents don't just answer questions. They increasingly take actions.
Lost card
Customer selects debit card •••• 4421. Was that exact card confirmed and frozen?
Transaction dispute
Customer disputes €129.50 at Merchant X. Did the agent open the dispute against that transaction — not another one?
Customer-data retrieval
Customer asks for transactions from Account A. Did the agent retrieve and disclose data only from Account A?
Advice and recommendations
Did informational guidance cross into personalised recommendation when the workflow did not permit it?
How does your organisation decide an agent release has met its approved controls?
We're working with a small number of teams deploying customer- or employee-facing agents in high-consequence workflows.