Agent Team template
A meta-agent workflow that generates red-team cases, runs regression suites, scores agent behavior, and blocks production rollout when a workflow violates its safety contract.
One-message setup · 3 specialists · Live after setup
Designed for
AI companies shipping agents that need repeatable evals, guardrail testing, and release confidence
Outcome
Catch behavior regressions before agent workflows reach production users.
Say what should be true when the work is done.
Codelit runs one labeled Sample with no account, model, or external App calls.
Review the evidence, then reuse the same brief with your data when it is useful.
Creates representative happy-path and adversarial tasks.
Reasoning model
Runs the suite and captures traces.
Deterministic execution route
Scores results and decides whether release can proceed.
Policy reasoning model
Loads the workflow goal, allowed actions, escalation policy, and output contract before the agent plans work.
A skill for generating scenario suites, red-team cases, pass/fail rubrics, and regression gates for agents.
Centralizes high-risk action checks for writes, secrets, customer data, billing, deploys, and public communications.
Exposes scenario datasets, trace reads, judge prompts, and release-gate results.
tools · resources · prompts · sampling
Generate or refresh scenarios from the workflow contract.
Execute scenarios and capture traces.
Compare metrics to thresholds and prepare a release decision.
Open it in Codelit and describe the result. Codelit handles common setup while Flow stays available for direct editing and Activity keeps the proof.
Use this teamTurns one scoped release request into a reviewed branch artifact, preview deployment, browser-verified result, pull request, production promotion, and Slack proof using the accounts the team already owns.
Correlates Sentry, Datadog, PagerDuty, and deployment evidence, proposes one bounded mitigation, pauses on the exact actions, and leaves status proof for responders.
Combines CRM context with bounded browser research, explains the qualification decision, creates an unsent email draft, and updates the lead after approval.