Navigate — Human-Agent Collaboration & Trust
The Navigate phase documents how agent-human interaction works in practice — not by default, but by explicit design per redesigned step.
Interaction modes by step
| Step | Mode | Rationale |
|---|---|---|
| 1. Normalize scope | Collaborative | Scope errors are cheap on paper, expensive in code; humans resolve contradictions |
| 2. Build and package | Collaborative | Engineering accountability stays human; agents accelerate parallel work |
| 3. Validation and triage | Agent-autonomous | CI/staging signals are largely mechanical; humans for ambiguity only |
| 4. Review and rework | Collaborative | Guardian pre-reviews; humans approve along risk thresholds |
| 5. Certify release readiness | Collaborative | Highest blast radius — human owns client-visible certification |
| 6. Merge and handoff | Collaborative | Agents produce runbooks; humans retain production authority |
Escalation principles
Escalate to human when:
- Inputs are incomplete, contradictory, or ambiguous
- Risk tier or reviewer routing cannot be inferred safely
- Client policy or contract interpretation is required
- Triage confidence is below threshold or blast radius is high
- Sensitive modules: security, data, billing, client-visible semantics
- Production promotion, gate waivers, or UAT acceptance ambiguity
Human override is always available: block merge/release regardless of agent recommendation; waive gates only through named approver path.
Building staff trust early
Introduce agents first on internal delivery mechanics — review packets, validation digests, runbook drafts, checklist assembly — where mistakes are cheap and clients are not exposed.
- Start with one repo or one squad
- Define success metrics: time-to-first-useful-review, revision rounds, CI noise reduction
- Weekly demos: engineers see agent output before it affects merge or release
- Label agents as assistants with bounded authority, not owners of production promises
Human override and correction
Every autonomous step produces a visible artifact (packet, digest, runbook draft) plus version or run ID for traceability.
Humans can:
- Reject or edit artifacts
- Block merge/release regardless of agent recommendation
- Waive gates only through named approver path
- Roll back using agent-drafted rollback notes — humans decide whether to execute
For implementation work, override means reopen/reassign the MR path — never silent patching without a record.
Feedback loops that improve agent behavior
Corrections become labeled feedback, not one-off complaints:
- Miscategorized risk → tagged entry tied to artifact ID
- Wrong reviewer route → ticket for routing rule update
- Missed flaky-test signature → baseline update
On a monthly cadence, review patterns and update rules, prompts, routing thresholds, and policy checks.
Measure improvement by: lower defect escape, fewer revision loops to approval, shorter time from failure to classified root cause — not "fewer escalations" alone.
Application order at Soft Pyramid
Apply internal delivery standards first (evidence packs, approvals, traceability, measurement). Roll out a client-facing playbook only once internal mechanics are trustworthy and repeatable.