Taso is an eval platform that tests model, prompt, tool, and MCP changes before production. It replays the same pinned scenarios with the current and proposed versions of your agent, then returns a deploy, review, or block decision with evidence.

What kinds of agent changes does Taso test?

Taso tests model swaps, prompt edits, tool and function changes, and MCP server updates. These changes can alter agent behavior without breaking traditional CI.

How does Taso fit into my CI/CD pipeline?

A Taso pilot connects your workflow to repeatable agent comparisons and release criteria. The website shows illustrative CLI, API, and CI integration examples; integration availability is confirmed during onboarding.

What does a Taso run measure?

Each run compares a baseline against a challenger across pinned scenarios. Taso reports quality, cost, latency, and reliability changes with trace evidence behind every decision.

Can Taso build evals from production traces?

Auto-Eval is a preview for early-access teams. It proposes checks from recurring agent behavior for your team to review. Only approved, reproducible cases belong in a pinned suite.

Taso · Test every agent change before production

Real behavior.
Better evals.

Turn production traces into evals that catch the next regression. Preview how Taso proposes checks from recurring behavior, ready for your team to review.

Proposed checkNeeds your review

Check the refund policy before acting.

Closed dispute → send for review, without issuing a refund.

Auto-Eval · early access. Combine traces with your team’s expertise and existing test suites.

Build a sandbox
for your agent.

Auto-Eval will draft synthetic environments and evals from your workflow requirements or traces. Review the setup, then test without touching real customer accounts.

Test data: Synthetic accounts and workflow state
Tools: Simulated responses and writes
Evals: Checks for the outcomes you expect

Explore the Auto-Eval preview

Support refund agent

Interactive sample · synthetic data

Hold this release

Block the unauthorized refund path

The new prompt fixes two failures, but issues a refund on a closed dispute. The policy violation blocks release even though the overall pass rate improves.

Change under testRefund prompt v12 → v13

BaselineFable 5.1tools-v4 · prompt-v12

ChallengerFable 5.1tools-v4 · prompt-v13

Configuration

Only the refund prompt changes. Model, tools, scenario inputs, and environment state stay fixed.

Suite: refund-policy-v1
Comparison: 6 scenarios · 1 illustrated attempt per version

Environment: Billing and dispute sandboxBilling records, dispute state, and refund policies.
Scenarios: 6 scenariosComplete sample catalog
Eval checks: 12 scenario checks10 deterministic · 2 judge rubrics

What movedBaseline Challenger

Scenarios passed: Baseline 3/6Challenger 4/6
Cost / run: Baseline $0.41Challenger $0.37
p95 latency: Baseline 1.18sChallenger 1.40s
New policy failures: Baseline 0Challenger 1

Cost and latency are illustrative.

Behind the decision

Scenarios & evidence

All 6 scenarios are included.

Recommended action

Fix the policy check before release.

Require the dispute-policy check before issuing a refund. Rerun the same suite and preserve the two improved cases. Keep the existing currency-conversion failure as a separate issue.

$ taso compare \ --workflow support-refund-agent \ --baseline "Fable 5.1 · tools-v4 · prompt-v12" \ --challenger "Fable 5.1 · tools-v4 · prompt-v13" \ --suite refund-policy-v1 \ --gate support_release_gate # Synthetic sample: 6 scenarios · 1 illustrated attempt per version ✓ cases passed: 3/6 → 4/6 ✓ 2 previously failing cases fixed ! 1 new policy regression: refund_policy_edge_case_17

Make your next agent release a measured decision.

In a 30-minute demo, we’ll compare two versions of your agent and walk through the evidence behind a release decision.

Auto-Eval (Preview): turn production traces into draft cases

Baseline and challenger runs in controlled environments

A change report with quality, cost, and latency evidence

Give your agent the right hill to climb

Real behavior.
Better evals.

Same conditions.
A different path.

Build a sandbox
for your agent.

See the change. Understand the risk.