The autonomy ladder
Six levels of autonomy, graded on code.
Carmakers grade self-driving from L1 to L5. We grade engineering teams the same way, step by step, from the first customer signal to the published release note. Here is the whole ladder.
- L1 AI engaged
- L2 AI activated Most enterprise squads
- L3 AI-native team
- L4 AI-native organization We take you here
- L5 Self-optimizing
- L6 Autonomous value network
Most enterprise squads
We take you here
The six levels
-
L1 · AI engaged
People carry every step. A copilot helps occasionally, inside one task.
No measurable change
-
L2 · AI activated
Everyone uses AI daily, with their own setup. The handoffs have not moved.
Most squads are here
Faster tasks, same lead time
-
L3 · AI-native team
Agents build and test on shared context. Humans authorize every merge and every release.
Capacity becomes measurable
-
L4 · AI-native organization
Agents run discovery and code production end to end, inside limits your people set. Humans keep goals, design judgment and the launch call.
We install this
Roles can be redesigned
-
L5 · Self-optimizing
The workflow improves itself. Agents test better ways of working against measured outcomes and keep only the wins.
Fewer human interventions
-
L6 · Autonomous value network
Agents coordinating work across company boundaries, with clients and suppliers.
A horizon, not a promise
The distance from L2 to L4 is not a bigger model or another license. It is an operating model, and it takes 30 days per squad to install.
The whole pipeline, all six levels
Fourteen steps carry a feature from a customer signal to a published release. This is who does each one, at each level.
- Human a person does it
- Human + agent a person directs, an agent drafts
- Agent runs without a person in the loop
- Gone the step no longer exists
| Step | L1 | L2 | L3 | L4 | L5 | L6 |
|---|---|---|---|---|---|---|
| 1 · Discovery | ||||||
| Signals | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Data and monitoring | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Priorities | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| 2 · Conception | ||||||
| Ticket | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Spec | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Design | Human | Human + agent | Human + agent | Human + agent | Human + agent | Human + agent |
| Handover | Human | Human + agent | Human + agent | Gone | Gone | Gone |
| 3 · Production | ||||||
| Code | Human | Human + agent | Agent | Agent | Agent | Agent |
| Test | Human | Human + agent | Agent | Agent | Agent | Agent |
| Review | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Deploy | Human | Human | Human + agent | Agent | Agent | Agent |
| 4 · Marketing | ||||||
| Documentation | Human | Human + agent | Human + agent | Agent | Agent | Agent |
| Launch assets | Human | Human + agent | Human + agent | Human + agent | Human + agent | Human + agent |
| Customer communication | Human | Human + agent | Human + agent | Human + agent | Human + agent | Human + agent |
Read the L4 column top to bottom. Discovery and production go dark: agents run them end to end. Design and customer communication stay mixed, because judgment and brand do not delegate. Handover disappears entirely, because a shared context removes the reason it existed.
One level at a time
Pick a level and read what a normal week looks like inside it.
L1 · AI engaged
People carry every step. A copilot helps occasionally, inside one task.
No measurable change · Squad: 8 dedicated people
1 · Discovery
Human
PM, PMM and CSM gather insight from calls, meetings, reviews and dashboards, then choose what matters.
-
Signals Human
PM, PMM and CSM run customer calls, sit in meetings, read store and support reviews, watch competitors, then write up what they heard.
-
Data and monitoring Human
Engineers and the PM read dashboards, usage reports and alerts, investigate, and run incident response.
-
Priorities Human
The PM weighs what the squad heard and measured, chooses the outcomes to pursue and owns the roadmap.
2 · Conception
Human
PM, PO and designer turn a chosen problem into tickets, specs and designs, then hand them to engineering.
-
Ticket Human
People collect feedback, triage bugs and write tickets.
-
Spec Human
The PM writes scope and acceptance criteria, with occasional copilot help.
-
Design Human
The designer defines flows, visuals and interaction. AI use is occasional.
-
Handover Human
People explain specs and designs in meetings and pass tickets between roles.
3 · Production
Human
Engineers write, test, review and release the change themselves.
-
Code Human
Engineers write code. Occasional copilot suggestions help locally.
-
Test Human
Engineers and QA define coverage, maintain tests and check results.
-
Review Human
Engineers inspect changes and decide whether to merge.
-
Deploy Human
Humans authorize releases and operate the CI/CD pipeline.
4 · Marketing
Human
A shared writer and PMM document the change, build the launch material and announce it.
-
Documentation Human
A shared technical writer collects changes from engineers and the PM, writes the docs and prepares release notes.
-
Launch assets Human
A shared PMM and the designer build demos, screenshots and sales or support material by hand for each launch.
-
Customer communication Human
A shared PMM works with the PM to brief sales and support, then announces the feature to the relevant customers.
L2 · AI activated
Everyone uses AI daily, with their own setup. The handoffs have not moved.
Faster tasks, same lead time · Squad: 8 dedicated people
1 · Discovery
Human + agent
Each role summarizes its own inputs with its own assistant. The PM still weighs the evidence and sets priorities.
-
Signals Human + agent
Each of them uses their own assistant to transcribe calls, summarize meetings and cluster reviews, then checks the summary against what was actually said.
-
Data and monitoring Human + agent
They hand AI selected funnels, logs and alerts to help investigate. Engineers confirm the cause and apply the fix.
-
Priorities Human + agent
The PM uses AI to compare options and size opportunities, then checks the evidence and owns the roadmap.
2 · Conception
Human + agent
Each role drafts with its own assistant. Meetings and handovers still carry the work between roles.
-
Ticket Human + agent
The PM turns selected feedback into ticket drafts with AI. Engineers structure bug reports the same way. The PM curates the backlog.
-
Spec Human + agent
The PM feeds context to an assistant to draft a PRD, scope and acceptance criteria, then resolves the gaps and approves.
-
Design Human + agent
Designers explore layouts, copy and prototypes with AI tooling, then direct the result and check usability.
-
Handover Human + agent
AI prepares handover notes. The meetings still happen, and context still moves between people.
3 · Production
Human + agent
Engineers use Claude Code or Cursor on selected tasks. Every merge and every release is human.
-
Code Human + agent
Engineers run Claude Code or Cursor with personal instructions on selected tasks, supplying context, steering the work and integrating the result.
-
Test Human + agent
They ask AI to draft tests, suggest edge cases and explain failures, then choose coverage and validate the results.
-
Review Human + agent
AI summarizes diffs and flags possible defects on selected PRs. Reviewing engineers verify the findings and approve every merge.
-
Deploy Human
Engineers check readiness, authorize the deployment, verify the rollout and decide whether to roll back. AI does not touch this.
4 · Marketing
Human + agent
Writer and PMM draft with AI. The PM confirms the claims and the PMM owns positioning.
-
Documentation Human + agent
The writer drafts docs from specs and code changes with AI. Engineers verify accuracy before publication.
-
Launch assets Human + agent
The PMM drafts launch copy with AI and repurposes it across formats, checks the claims, and the designer keeps the visuals on brand.
-
Customer communication Human + agent
The PMM drafts briefs, demos and announcements with AI. The PM confirms the claims, the PMM chooses audience, timing and channel.
L3 · AI-native team
Agents build and test on shared context. Humans authorize every merge and every release.
Capacity becomes measurable · Squad: 8 dedicated people
1 · Discovery
Human + agent
Shared agents read feedback, usage and telemetry continuously. The squad chooses the problems to solve.
-
Signals Human + agent
Shared agents transcribe, summarize, cluster and track competitors into one picture. The squad checks it against its own field knowledge.
-
Data and monitoring Human + agent
Agents read telemetry and product usage, propose triage and surface trends. Humans authorize remediation.
-
Priorities Human + agent
Agents rank opportunities against outcome data. Humans choose the problems to solve.
2 · Conception
Human + agent
Agents draft tickets, specs and prototypes from shared context. Humans validate intent and experience.
-
Ticket Human + agent
An agent drafts tickets from feedback and telemetry. The PM curates.
-
Spec Human + agent
An agent drafts specs from team context. The PM validates intent.
-
Design Human + agent
An agent proposes flows and prototypes. The designer validates the experience.
-
Handover Human + agent
Agents package executable specs. Humans resolve the gaps between roles.
3 · Production
Human + agent
Agents implement and test scoped tickets. Humans authorize every merge and every release.
-
Code Agent
Background agents implement scoped tickets and open PRs. Humans review at the next gate.
-
Test Agent
Agents generate and run tests against agreed acceptance criteria. Failures return to the coding agent.
-
Review Human + agent
Agents pre-review PRs. Humans resolve findings and authorize every merge.
-
Deploy Human + agent
Agents assess readiness and prepare the rollout. A human authorizes the release.
4 · Marketing
Human + agent
Agents draft docs and launch material from the approved spec. Humans review the messaging.
-
Documentation Human + agent
Agents draft docs alongside the feature and check examples and links. Engineers and the writer review before release.
-
Launch assets Human + agent
Agents assemble demo scripts, screenshots and enablement material from the shipped feature. The PMM and designer approve before use.
-
Customer communication Human + agent
Agents prepare launch material from the approved spec and release notes. The PMM reviews messaging and coordinates the rollout.
L4 · AI-native organization
Agents run discovery and code production end to end, inside limits your people set. Humans keep goals, design judgment and the launch call.
Roles can be redesigned · Squad: 6 dedicated people
1 · Discovery
Agent
Agents watch every signal, rank the opportunities and start work within goals your people set.
-
Signals Agent
Agents collect and synthesize every customer, market and internal signal continuously, and flag what changed since the last read.
-
Data and monitoring Agent
Agents detect incidents, roll back or repair, verify recovery, and feed usage trends and post-incident findings into the next priorities.
-
Priorities Agent
Agents rank opportunities and choose work within human-set goals. Anything outside those goals escalates.
2 · Conception
Human + agent
Agents turn goals into executable specs and the separate handover disappears. Design keeps human judgment.
-
Ticket Agent
Agents capture, deduplicate, prioritize and route work across squads automatically.
-
Spec Agent
Agents generate executable specs and acceptance criteria from shared context. Ambiguity escalates.
-
Design Human + agent
Agents generate flows and device variants. Humans judge novel experiences and hold the design standard.
-
Handover Gone
Shared context and executable contracts remove the separate manual handoff.
3 · Production
Agent
Agents implement, test, review and release passing changes, restoring the known-good version on failure.
-
Code Agent
Agents implement and integrate changes across repos, escalating only outside delegated scope.
-
Test Agent
Agents run regression, security and device checks on every change, and block failures.
-
Review Agent
Independent agent review and automated checks gate every merge. Regressions trigger rollback.
-
Deploy Agent
Agents release passing changes behind flags, monitor the rollout and automatically restore the known-good version on failure.
4 · Marketing
Human + agent
Agents publish docs and prepare launch material with each release. Humans approve brand and positioning.
-
Documentation Agent
Agents maintain guides, API docs and release notes with each change, validate the examples and publish with the release. Product engineers resolve unclear behavior.
-
Launch assets Human + agent
Agents produce launch assets per audience and per device from the shipped feature. Humans approve brand and claims.
-
Customer communication Human + agent
Agents prepare internal briefs and customer announcements from the shipped feature. Product engineers verify behavior, the shared PMM approves positioning and timing.
L5 · Self-optimizing
The workflow improves itself. Agents test better ways of working against measured outcomes and keep only the wins.
Fewer human interventions · Squad: ~4 product engineers
1 · Discovery
Agent
Agents check predicted value against delivered outcomes and improve how they find and rank work.
-
Signals Agent
Agents measure which sources actually predicted value, drop the noisy ones and test new ones.
-
Data and monitoring Agent
Agents replay incidents to test detectors, runbooks and metrics, trial better versions and revert regressions.
-
Priorities Agent
Agents compare predicted value with delivered outcomes, test new ranking methods and keep the winners.
2 · Conception
Human + agent
Agents trace rework back to spec and design gaps, and test better methods against it.
-
Ticket Agent
Agents measure missed or duplicate work, test intake and routing changes, and deploy better rules.
-
Spec Agent
Agents trace rework to spec gaps, evaluate new templates on past tasks, and promote those that reduce failures.
-
Design Human + agent
Agents test design variants against usage and accessibility outcomes. Humans approve changes to the design system.
-
Handover Gone
Agents improve context routing from failed dependencies. The manual handoff stays removed.
3 · Production
Agent
Agents trial better prompts, tools, coverage and rollout policies, keeping only measured wins.
-
Code Agent
Agents replay failed tasks with revised prompts, tools or routing. Successful candidates enter a controlled trial.
-
Test Agent
Agents turn escaped bugs into regression cases and evaluate improved coverage against protected checks.
-
Review Agent
Agents measure missed defects and false alarms, test reviewer changes, and promote gains without weakening required checks.
-
Deploy Agent
Agents trial new rollout policies against recovery time and failure rate, promoting wins or restoring the prior policy.
4 · Marketing
Human + agent
Agents close documentation and asset gaps from real support and usage evidence. Humans keep major launches.
-
Documentation Agent
Agents use support questions and failed searches to find documentation gaps, test improvements and keep docs aligned with the shipped product.
-
Launch assets Human + agent
Agents measure which assets convert or deflect support tickets and keep the formats that work. Humans still own the brand.
-
Customer communication Human + agent
Product engineers own routine feature communication, with agents drafting and distributing inside agreed rules. Shared PMM expertise supports major launches.
L6 · Autonomous value network
Agents coordinating work across company boundaries, with clients and suppliers.
A horizon, not a promise · Squad: ~4 product engineers
1 · Discovery
Agent
Could surface and rank shared opportunities with client and partner agents.
-
Signals Agent
Could pool signals with client and partner agents under agreed data terms.
-
Data and monitoring Agent
Could coordinate detection, recovery and shared usage measurement across client and supplier agents.
-
Priorities Agent
Could rank and commit to shared opportunities with client and partner agents.
2 · Conception
Human + agent
Could agree executable specs across companies within delegated commercial terms.
-
Ticket Agent
Could exchange and route work across companies through agreed intake contracts.
-
Spec Agent
Could align executable specs across companies within delegated commercial terms.
-
Design Human + agent
Could adapt experiences across client brands. Taste and brand decisions stay human.
-
Handover Gone
Could replace cross-company handoffs with shared machine-readable contracts.
3 · Production
Agent
Could build, verify and release jointly with agents in client and supplier systems.
-
Code Agent
Could build and integrate changes jointly with agents in client and supplier systems.
-
Test Agent
Could run shared compatibility and contract tests across participating platforms.
-
Review Agent
Could verify cross-company changes using shared evidence and delegated approval policies.
-
Deploy Agent
Could coordinate staged releases and recovery across partners that support the required controls.
4 · Marketing
Human + agent
Could coordinate documentation and launches across partner products under shared rules.
-
Documentation Agent
Could synchronize documentation and release notes across client and partner products.
-
Launch assets Human + agent
Could produce co-branded launch material with partner agents. Brand and commercial judgment stay human.
-
Customer communication Human + agent
Could extend product-engineer ownership to feature marketing, with agents coordinating partner launches.
What actually moves a squad up
L2 → L3
Stop optimizing individuals. Put the context the agents need into the repo, make the quality gates machine-checkable, and give the squad one shared agent setup instead of nine personal ones.
L3 → L4
Move the human gate from every merge to the boundaries of a delegated scope. That requires tested rollback, independent agent review, and a written answer to the question of what an agent may decide alone.
L4 → L5
Instrument the workflow itself. Once you measure which prompts, tests and rollout policies produce outcomes, agents can run the improvement loop and you keep only what wins.
Where does your squad sit?
Twelve questions, one per step in the pipeline. You get your level, a score per phase, and the three steps holding you back.
Free, about four minutes, no email required to see your result.
Or read what L4 costs.
We embed in one squad for 30 days and leave you with a measured before and after. Three engagements, published prices.