Stonepath LabsView demo

TAQ · RELEASE CONTROL FOR AI AGENTS

Commerce Agent v1.3release gate

TAQ

Turn failed AI agent runs into release gates.

Save real agent failures as regression tests, replay them before prompt, model, tool, or workflow changes ship, and block releases that repeat old mistakes.

Open-source capture layer·replayd
Commerce Agentv1.3Release candidate

18 known failures replaying

Checking release candidate
release gatelive
Passed
17
Failed
1
Gate
BLOCKED
Critical failure

image request routed to text response

Input
"Can I see a picture?"
Expected
send_product_image
Actual
ai_agent_text_only
Trace
Captured
Saved as test
Replayed
Blocked
DecisionBLOCKED
release stopped before deploy

TAQ RELEASE LOOP · REPEATABLE PER AGENT VERSION

every failed run gets a path to the gate

1failed run

real behavior captured

2saved regression test

failure becomes a case

3replay before release

known failures replay

4approve/block

gate makes the call

The tension

Agents are moving into production. Their release process has not caught up.

AI agents no longer just answer questions. They call tools, update records, issue refunds, retrieve documents, route workflows, and touch customer or financial data.

Normal software ships with regression tests. Agent deployments often ship with manual checks, screenshots, eval spreadsheets, intuition, and incident memory.

A prompt fix that solves one failure today can silently reintroduce another failure tomorrow.

The gate starts with the failures a team already knows.

incident memory

old failures waiting for a gate

repeated case
01
failure returned after model change
Agent failures do not disappear after one prompt fix. They return when models change, retrieval shifts, tool definitions change, or new workflow logic ships.
02
no standard release gate
Software teams have CI pipelines and regression suites. Agent teams often rely on manual review, intuition, and production incidents.
03
real failures are the test cases
The clearest signal of what an agent must not do already exists in failed runs, logs, tool traces, and user complaints.

Text evals are not enough

The model can sound right while the system does the wrong thing.

A customer asks to see a product image. The agent replies with a reasonable text description. But the workflow should have sent an image.

A text evaluator might pass the answer. A release gate should block the behavior.

TAQ checks the system action, not just the language.

intent · route · tool call · output payload · policy decision

Text eval

Checks whether the answer sounds correct.

Example
Product description is relevant and fluent.
Result
May pass.

TAQ release gate

Checks whether the workflow did the right thing.

Example
Expected route: send_product_image
Actual route: ai_agent_text_only
Result
BLOCKED

WHAT TAQ CHECKS

TAQ checks behavior, not just answers.

Replay cases can assert the route taken, tool called, output shape, retrieved context, policy decision, or semantic expectation.

Select assertion surface
Behavior inspectorroutes

commerce_image_request_001 · replay case

BLOCKED
Failed old case

image request routed to text response

Gate verdict
Release decisionBLOCKED
Expected
send_product_image
Actual
ai_agent_text_only
replay pathcaptured → asserted → gated
failed runreplay caseblocked

OPEN-SOURCE CORE

replayd turns failed runs into replayable fixtures.

Use replayd to capture the failure, define what should have happened, and replay it before future prompt, model, tool, retrieval, or workflow changes ship.

pip install replayd
replaydtest runner
replayd run commerce_image_request_001
FAIL
FixtureReplay outputGate decision
case_idcommerce_image_request_001
input"Can I see a picture?"
expected_routesend_product_image
actual_routeai_agent_text_only
Assertions
route_taken == send_product_image
final_output.type == media
image_url exists
Replay output
FAILtrace.router.route_taken
expectedsend_product_image
actualai_agent_text_only
Gate footer
release_decisionBLOCK

The failed route becomes a replayable case. TAQ can use the same decision path as a release gate.

HOW TAQ WORKS

From failure to release decision.

The same known failure moves through capture, expectation, replay, and a release gate. Each stage removes ambiguity from the next ship.

01
failed run

Capture the failure

Log the full run context: input, output, tool calls, retrieved context, prompt version, model, and what went wrong.

Stage artifact
inputtool_callscontextmodel
02
replay fixture

Define expected behavior

Mark what should have happened and what must not happen again.

Stage artifact
expected: send_product_imageblocked: ai_agent_text_only
03
replay suite

Replay against the next version

Run the saved case against a new prompt, model, retrieval, or tool configuration before production.

Stage artifact
prompt v1.3model: production-smallretrieval@v2
04
release gate

Approve or block the release

TAQ returns a gate decision. Every failure shows exactly what broke.

Stage artifact
APPROVEDBLOCKEDESCALATED

RELEASE GATE

Watch TAQ block a repeated agent failure.

A release candidate enters the gate. Known failures replay. If one repeats, TAQ blocks the release before it reaches users, data, tools, or money.

release evaluation running
Commerce Agentv1.3Release candidate

18 known failures replayed

Checking
replay railevaluating
Passed
17
Failed
1
Gate
BLOCKED
Critical failure

image request routed to text response

Input
"Can I see a picture?"
Expected
send_product_image
Actual
ai_agent_text_only
Trace
Captured
Saved as test
Replayed
Blocked

Release decision

DecisionBLOCKED
ReasonKnown failure repeated
OutcomeReview before deploy

The old failure is caught before release. The candidate does not ship until the behavior is reviewed.

Have a failure this should catch?

PRICING

Start locally with replayd. Add TAQ release gates when agents touch users, tools, data, or money.

Developer

Free / Open source

Start with replayd locally.

Start locally with replayd. Capture failed runs and turn them into replayable fixtures in your repo.

  • local replay fixtures
  • deterministic assertions
  • tool, route, and output checks
  • commit cases to your repo
View GitHub

serious team path

Team Beta

Private beta

Release gates for real agents.

For teams shipping AI agents or client automations that need replay suites and release-blocking decisions before changes go live.

  • shared replay suites
  • release-blocking assertions
  • prompt/model/tool regression checks
  • workflow review before deploy
  • early product support
Send a failed run

Enterprise

Future / Contact

Future governance and control.

For organizations that need governance around agents touching customers, data, tools, or money.

  • approvals and audit trails
  • team permissions
  • release history
  • compliance-ready reporting
  • policy gates
Contact us

CONTACT

Show us one failure you never want users to see again.

Send us one real failure: what the user asked, what the agent or workflow did, and what should have happened instead. We'll help turn it into a replay test.

failed runreplay caserelease gate
hello@stonepathlab.net