Story 03 · Governed multi-agent operating model

Eight roles. Shared context. Autonomy only where the gate exists.

This is the AI-SDLC operating system: specialist agents with responsibilities, boundaries, model selection, handoffs, and review gates. The DevOps engineer is a role in the system, not the identity of the work.

General assistants collapse every job into one prompt. Quality then depends on whoever typed. A delivery system needs named roles, a shared brief, and a place where a human must still say yes.

Planner

Frames intent, constraints, and the sequence of work. Does not invent architecture or write production code.

Model: stronger reasoning. Handoff: structured brief to architect. Gate: human confirms scope before design starts. Autonomy limited to draft plans.

Architect

Turns the brief into options, interfaces, and risks. Does not implement the chosen option by default.

Model: long-context reasoning. Handoff: design record to implementer and reviewer. Gate: human approves the option. Autonomy limited to options, not decisions.

Implementer

Produces the change inside the approved design. Does not merge, release, or redefine scope.

Model: code-specialist. Handoff: patch plus notes to tester and reviewer. Gate: cannot skip tests or review. Autonomy limited to the approved unit of work.

Tester

Designs and runs checks against the stated behaviour. Does not waive failures to keep the flow moving.

Model: precise, low-creativity. Handoff: evidence pack to reviewer. Gate: failing checks block progress. Autonomy limited to test design and execution.

Reviewer

Judges fitness against the brief, standards, and risk. Does not rewrite the change as a second implementer.

Model: critical, citation-friendly. Handoff: approve, request change, or escalate. Gate: human still owns merge. Autonomy limited to recommendations.

Security auditor

Looks for the failure modes assistants introduce: secrets, unsafe defaults, dependency invention, data leakage.

Model: security-tuned. Handoff: findings with severity. Gate: high-severity blocks release. Autonomy limited to detection, never exception-granting.

DevOps engineer

Integrates the change into the real delivery path: build, environment, release evidence. This is the technical foundation that lets AI enter production systems.

Model: systems-oriented. Handoff: pipeline and runtime evidence. Gate: humans own production change. Autonomy limited to proposed automation, not live production action.

Documentation

Records decisions, interfaces, and operating notes so the next agent and the next human share the same context.

Model: concise long-context. Handoff: living brief back into the planner. Gate: humans accept the record as the source of truth. Autonomy limited to drafts.

Shared context and review gates

Every role reads the same brief: intent, constraints, decisions already taken, and what a human has already refused. Handoffs are artefacts, not chat. Review gates sit after architecture, after implementation, after security, and before release. Autonomy rises only when intervention frequency and quality outcomes say it is safe.

This page describes the operating system. It is not a product catalogue and not an unsupervised agent platform.

See the AI-SDLC Audit Back to enterprise adoption