Replay the failure before your AI agent ships it.
Traceforge turns live agent traces into deterministic eval suites, policy checks, and release decisions for teams building customer-facing AI workflows.
Release decision
A release path built from actual traces.
Traceforge keeps the launch decision close to engineering reality: capture what the agent tried, replay the risky branches, score policy and tool behavior, then attach evidence to the release.
SDK records prompts, tool calls, retrieved context, latency, cost, and user-safe redactions from production runs.
Flaky agent paths are frozen with deterministic mocks for APIs, documents, and retrieval snapshots.
Policy checks measure refusal quality, source grounding, escalation accuracy, and action reversibility.
Merge checks block risky releases and produce audit-ready evidence for product, security, and support.
Proof a reviewer can act on.
The site foregrounds concrete product artifacts instead of vague AI claims: scenario traces, thresholds, evaluator names, and a decision table that mirrors a real launch review.
| Evaluator | Signal | Release rule | Current state |
|---|---|---|---|
| Policy guard | Detects unauthorized refund, claim payout, or external-email action. | Block if severity is high. | selected state visible |
| Grounding check | Compares final answer to approved source snippets and tool responses. | Require 90% source support. | hover rows highlight |
| Escalation check | Confirms agent hands off when confidence, policy, or missing-data limits are hit. | Require handoff within two turns. | error state shown |
Structured design brief embedded in the product story.
This prototype uses the prompt as a senior-design-ready brief: purpose, layout, visual system, interaction states, responsive behavior, and implementation constraints are expressed through the page itself.
Purpose
Help AI product and platform teams decide whether an agent release is safe enough to ship using replayable evidence.
Layout
First viewport pairs a launch promise with an interactive eval artifact. Later sections move from architecture to proof to constraints.
Visual system
Warm paper, near-black code surfaces, quiet grid lines, and one green accent create a technical but not sterile tone.
States
- Selected scenario tabs
- Hover navigation and row affordances
- Disabled export until evidence exists
- Loading replay spinner
- Success and error release decisions
Responsive behavior
Desktop presents copy and artifact side by side. Mobile stacks the artifact below the promise, keeps the run action visible, and turns tables into horizontal inspection surfaces.
Constraints
No external app shell, no generic stock imagery, no decorative AI glow. The product artifact carries the concept and stays usable without a backend.
Put an eval gate in front of your next agent release.
Upload a failed trace, pick the policy threshold, and generate evidence your reviewer can verify.