Capability 03 of 04 · Organization
Team Structure & Roles
How teams are organized and what roles exist to support AI-augmented engineering.
5
maturity levels
15
practices
15
guides
L1 · Stage 01Assisted
Practices - what it looks like3 guides
- Traditional roles: dev, QA, PMDev writes, QA tests, PM specifies. Clear scopes and clean handoffs, and the model that predicts exactly where AI adoption will create pressure.guide→
- Seniors review and fix AI-generated code; human-skill preservation (reject code you can't understand even if it works)The anti-pattern where juniors generate code faster than they can verify it, and seniors spend their days debugging the results instead of building.guide→
- AI being evaluated in the team's stack"AI doesn't work in our environment" is the most common organizational statement that blocks progress at L1.guide→
Criteria - what to measure
- 01The team has standard engineering roles (developer, QA, PM)
- 02Senior developers review and fix AI-generated code
- 03Team is open to experimenting with AI-assisted workflows
- 04At least one person informally champions AI tool usage
L2 · Stage 02Delegated
Practices - what it looks like3 guides
- A named AI champion per team, with time actually allocated to the roleEach team has one named AI champion with time actually allocated to the role, rather than an unfunded expectation resting on whoever cared most.guide→
- Context engineer role (initial)Making the codebase legible to agents, through CLAUDE.md files, MCP integrations and machine-readable docs, becomes somebody's actual job.guide→
- Training: how to instruct confidently and then verify confidently, which is not the same as reading every lineTraining people to instruct confidently and then verify confidently, which is not the same skill as reading every line the agent wrote.guide→
Criteria - what to measure
- 01AI champion is designated per team with allocated time (not just informal interest)
- 02Developers have been trained on how to give an agent a task
- 03Context engineer role exists (initial, possibly part-time) for maintaining agent instruction files
- 04Champion has a regular cadence for sharing learnings across the team
- 05Training materials are documented and available for new hires
L3 · Stage 03Systematic
Practices - what it looks like3 guides
- Review shifts up the lifecycle: judgment relocates rather than disappears - problem selection, architecture, the quality bar, which signals to trust, and shipping authority stay human even when authorship does notWhen most code in a PR is agent-generated, judgment does not disappear - it relocates to problem selection, architecture, the quality bar, which signals to trust, and shipping authority.guide→
- Platform Engineer (AI tooling); Harness Engineer as the consolidated named skill - the harness, not the model, is the asset that survives a vendor swapThe Platform Engineer specializing in AI tooling owns the harness - the asset that survives a vendor swap, where the model does not.guide→
- Context Engineer = full role (now mainstream - dedicated job postings across industry)At L3, context engineering graduates from "something the champion does in their spare time" to a full-time engineering role with its own scope, career path, and organizational standing.guide→
Criteria - what to measure
- 01Team's primary activity has shifted from writing code to evaluating and reviewing AI-generated code
- 02Platform Engineer role with AI tooling responsibility exists on the platform team
- 03Context Engineer is a full dedicated role (not part-time, not combined with other duties)
- 04Role definitions are updated to reflect AI-augmented responsibilities
- 05Hiring criteria include AI tool proficiency
L4 · Stage 04GovernedMost teams aim here
Practices - what it looks like3 guides
- Span of control = how many agents you can effectively supervise; the binding limit is the orchestrator's context, not tokens - batches capped at 2-4, status polling restricted, overlapping file ownership read as a signal to consolidateHow many agents one developer can actually supervise at once. The binding limit is the orchestrator's context, not the token bill, which caps a batch at two to four.guide→
- Developer = manager of agent fleet (now a product default: Cursor Run Mode, Claude agent view, Antigravity, MultiDevin)At L4, the developer's primary job is not to write code - it's to manage a fleet of AI agents that write code.guide→
- Each running agent has a visible health state, so a stalled or drifting one is noticed without being hunted forSteve Yegge's "keep your Tamagotchi alive" framing captures a crucial insight about working with AI agents at L4: they are not fire-and-forget automations.guide→
Criteria - what to measure
- 01Span of control is measured: how many parallel agents each developer effectively supervises
- 02Performance evaluation includes agent supervision effectiveness (not just personal code output)
- 03Developer role is formally defined as "manager of agent fleet"
- 04A span-of-control limit is defined per role and derived from what the orchestrator can actually keep in context
- 05Agent supervision training is part of standard developer onboarding
L5 · Stage 05Self-improving
Practices - what it looks like3 guides
- Agentic Engineer: orchestration + supervision + architectureThe Agentic Engineer is the L5 role that emerges when AI agents become the primary development modality and the human's job is to architect, orchestrate, and supervise rather than implement.guide→
- PEV loop: Plan → Execute → VerifyThe PEV loop - Plan, Execute, Verify - is the fundamental operating model for working with AI agents at high maturity.guide→
- Non-coder contributors via agent interfacesProduct managers, designers and domain experts direct agents to change the software directly, without writing code or waiting on an engineer.guide→
Criteria - what to measure
- 01PEV (Plan, Execute, Verify) loop is the standard workflow for all engineering tasks
- 02Non-coder contributors can produce software changes via agent interfaces
- 03Agentic Engineer role combines orchestration, supervision, and architecture responsibilities
- 04Agentic Engineer career ladder exists with defined progression criteria
- 05Non-coder contribution rate is tracked as an organizational capability metric
From the Field
Recent releases, projects and discussions the AI Radar classified into this capability.
discussion
I built a physical Tamagotchi that feeds on my Claude Code activity
article
Cybernetics and the “human-on-the-loop” in agentic coding
discovered
A virtual pet companion for your AI — Designed to provide in-context code review feedback with personality. Grow with your buddy and level up together. Works with Clau
article
Agentic Workforce Framework, an operating model for autonomous agent teams
article
GM just laid off IT workers to hire those with stronger AI skills
article
Thoughts on GitLab's workforce reduction" and "structural and strategic decisions"
article
AI didn't kill your junior pipeline. You did
article
Your Future job will be to keep AI on task
Which level is your team at in Team Structure & Roles?
The criteria above are what we check in a facilitated assessment. Run it yourself to place this capability, see which gates you have passed, and compare it against the other 3 in Organization.
Start the assessment →