pip install agentqa-sdk · live

Test your AI agent before it breaks a real workflow.

Record every run, catch failures with reasons, compare prompts and models — and know your success rate before you ship.

suite-run · sales_research_agent · prompt-v4
Success rate62%

−9 pts vs prompt v3

Top failureHallucinated revenue

×5

Cost / successful run$0.38
Worst test typesimilar-names

33% pass over 6 cases

The diagnosis your team can't produce by watching runs by hand.

1/3tasks failed by agents on structured benchmarks
12→66%agent task success in one year — capability outpacing trust
15minfrom signup to your first recorded trace
01 — record

One context manager. Every run recorded.

Wrap any Python agent — LangChain, browser-use, or plain code — and AgentQA captures the trace: steps, tool calls, screenshots, cost, latency. Define test cases with expected behavior and forbidden actions, then run the suite on every change.

  • 02Deterministic rules score what rules can score — forbidden tools, budgets, formats
  • 03An LLM judge handles policy adherence and unsupported claims
  • 04Uncertain cases go to human review — never guessed
python
import agentqa

with agentqa.track_run(
    agent_name="sales_research_agent",
    task_input=task,
) as run:
    result = my_agent.run(task)
    run.set_output(result)

# every step, tool call, screenshot,
# cost and verdict — recorded

Know your success rate, not your vibe

Run your agent against a real test suite and get pass/fail per case — with the failure reason attached, from hallucinated facts to wrong tool calls.

Catch regressions before customers do

Every prompt tweak and model update runs against the same suite. See exactly what got better, what broke, and what it costs per successful task.

Replay every failure

Full traces — steps, tool calls, screenshots — so “what happened?” takes one click, not an afternoon of archaeology.

Verdicts with reasons, not just logs

Observability tools show you the trace. AgentQA rules on whether the job was done correctly — and classifies every failure against a library of real agent failure patterns.

Passed Failed Needs review Risky
hallucinated informationwrong record selectedfake citationtool looppolicy violationpremature completionunauthorized actionexcessive costignored constraintsensitive data risk+8 more

Simple pricing

Start free, upgrade when your test suites do. Payments are handled by Polar — taxes included, cancel anytime.

Free

$0

Kick the tires on one agent.

  • 1 project
  • 50 judge evaluations / day
  • Full trace capture
  • Shareable reports
Start free

Starter

Recommended

$99/mo

For a team shipping its first agents.

  • 3 projects
  • 500 judge evaluations / day
  • Everything in Free
  • Email support
Start with Starter

Team

$499/mo

For agencies running many workflows.

  • 10 projects
  • 2,000 judge evaluations / day
  • Everything in Starter
  • Priority support
Start with Team

Agent reliability audit

Shipping agents for clients? We test one agent workflow end to end — 20+ cases, failure classification, and a shareable report with hard before/after numbers. $1,000–$5,000, one to two weeks.

Book an audit

Test your agents like software.

Get started — first trace in 15 minutes