01
Define
voice:
formality: low
warmth: very-high
directness: high
humor:
target: low
style: drySkip to contentExplore in Playground 
Governance SDK for AI Agent Behavior
Govern agent behavior across your fleet. Start in 90 seconds.
Define voice and behavioral policy as composable YAML profiles. Validate safety at build time. Compile model-aware system prompts. Evaluate adherence across your agent fleet.
Define. Validate. Compile. Evaluate.
02
Validate
$ traits validate my-profile.yaml
S001 PASS No unsafe instructions
S008 PASS Grounding constraints present
Schema v1.6 valid — 0 errors, 0 warnings03
Compile
[VOICE TARGETS]
formality: low
warmth: very-high
directness: high
[CAPABILITY BOUNDARIES]
Tools: (none — advisory only)
Constraints:
- Never claim actions without tool confirmation04
Evaluate
$ traits eval my-profile.yaml --tier 1
Tier 1 — 5 samples
avg score: 0.72
vocabulary: 4/5 preferred, 0 forbidden
dimensions: directness 0.86, warmth 0.71
PASS — all checks above thresholdWhy governance, not just prompting
Composable Policy
Define voice, rules, and capability boundaries as structured YAML. Extend base profiles with array-merge composition. Share constraints across your agent fleet without copy-pasting prompt fragments.
Build-Time Safety
Validate profiles against 8 safety checks before they reach production. Catch unsafe instructions, missing grounding constraints, and schema violations at CI time — not after deployment.
Fleet Evaluation
Score agent responses against profile policy with three evaluation tiers. Tier 1 runs locally in milliseconds. Tier 2 uses embeddings. Tier 3 uses LLM judges. Gate releases on adherence scores.
Starter policy profiles
Each profile defines voice targets, behavioral rules, and capability boundaries for a specific domain. Use them as base policies or extend them for your fleet.
Healthcare Companion
High-empathy care navigation with medical escalation rules and grounding constraints that prevent unsupervised clinical claims.
Scenario
I've been having chest pains after exercise. Should I be worried?
Response style
Acknowledge fear clearly, escalate urgent symptoms immediately, and frame next steps as recommendations for the care team.
Voice dimensions
Integration is two lines
const compiled = compileProfile("profiles/resolve.yaml", { model: "gpt-4o" });
const response = await openai.chat.completions.create({
model: "gpt-4o",
messages: [
{ role: "system", content: compiled.text },
{ role: "user", content: userMessage }
]
});Trust signals