Has your agent earned the right to do this work?
TrustPlane is the authority layer above AI coding agents. It classifies risk, authorizes runtimes, scores trust and produces a tamper-evident audit trail for every governed task across every runtime.
Agent runtimes answer: can the agent do the work? TrustPlane answers: has the agent earned authority to do it?
Requires Node.js 18+ and PostgreSQL.
git clone https://github.com/rayyagari2-create/agentic-workforce-framework
cd agentic-workforce-framework
cp .env.example .env
npm install
npm run demo:setup
npm run demoThe demo classifies 5 tasks by risk, fires an approval gate on high-risk items, scores agent performance on D1-D4 dimensions and writes a hash-chained audit log.
TrustPlane blocks, authorizes or supervises a runtime based on earned trust, not capability.
The Sprint 3 authorization demo is deterministic and example-driven. Production runtime authorization uses the same schemas and audit shape.
npm run authorize:blocked
npm run authorize:authorized
npm run authorize:supervisedExample output:
Decision: BLOCKED
Runtime: codex
Task class: db_migration
Risk lane: critical
Required tier: HIGH
Current tier: PROVISIONAL
Reason:
Codex has not earned sufficient trust for critical lane
database migrations in this workspace.
Recommended: claude_code
Every task goes through five stages:
- Classify: task class and risk lane from five dimensions
- Authorize: check runtime trust tier against required tier
- Execute: runtime runs under approved scope and controls
- Score: D1-D4 trust scoring from real artifacts
- Verify: hash-chained audit trail verified end to end
Trust tiers: PROVISIONAL, RESTRICTED, STANDARD, HIGH Scoring dimensions: D1 Correctness, D2 Observability, D3 Policy, D4 Recurrence
Sprint 3 introduces the named capability: Trust-Based Runtime Authorization.
Not routing. Authorization.
AWF evaluates agent and runtime behavior across D1-D4 scoring dimensions, then assigns a task-specific trust tier that determines whether the runtime is authorized, supervised or blocked.
| Runtime | UI refactor | DB migration | Security fix |
|---|---|---|---|
| Cursor | HIGH | PROVISIONAL | PROVISIONAL |
| Codex | STANDARD | PROVISIONAL | RESTRICTED |
| Claude Code | STANDARD | STANDARD | HIGH |
Sprint 3 public deliverables: four new schemas, nine worked examples, four architecture docs, and the authorize-task CLI demo above.
Sprint 2 proved cross-runtime governance across Claude Code and Codex from one control plane. Sanitized replay in examples/cross-runtime/.
schemas/v1/ JSON schemas for all governance artifacts
examples/ Worked examples for authorization decisions,
trust profiles and risk profiles
docs/ Architecture, D1-D4 scoring, task risk
profiling, execution substrates
agents/ Reference agent instruction files
calibration/ D1-D4 rubric and scoring anchors
hooks/ OS-level hook enforcement examples
database/migrations/ Postgres governance schema and audit log
services/ Audit service, scorer, governance engine
TrustPlane does not spawn agents, generate code, route subagents or capture tool-call telemetry. Those belong to execution substrates like Claude Code, Codex, Cursor, Devin and Maggy.
TrustPlane sits above them and decides who is authorized to run what and under which controls.
See docs/execution-substrates.md for the full three-layer architecture.
From the private reference implementation:
- 120+ governed sessions scored
- Zero successful hook bypasses
- Claude Code and Codex adapters proven at Gate 3
- Cross-runtime demo verified. One audit chain, two runtimes
Metrics are self-reported and not independently audited.
AgentTaskManifest Task contract before dispatch
QAVerdict Structured verdict with per-criterion evidence
FailureRecord 17-class failure taxonomy with recurrence
TrustScore D1-D4 session score and trust tier
TrustSubject Accountable identity AWF scores
TaskRiskProfile 5-dimension risk scoring
TrustCapabilityProfile Earned trust per runtime per task class
RuntimeAuthorizationDecision Authorization output with audit event
AuditEvent Canonical per-event audit shape with user_id
All schemas in schemas/v1/ are AJV Draft 2020-12 compatible.
npx agentic-workforce-framework@latest initScaffolds AWF governance into any repo.
Schema extensions, case studies and documentation corrections welcome. See CONTRIBUTING.md.
MIT. See LICENSE.
Originated by Ramesh Ayyagari, 2026