Six levels of factory autonomy, the gates between them, and our forecast for how far 2026 gets us. Written down now so you can grade us later.
Predictions about AI are cheap because nobody checks them. This page is a rubric, not a vibe: every gate below is either true or false of a real team shipping real software. We update it as gates open, and we do not move the goalposts.
Borrowed from how self-driving cars are graded, applied to how software gets built. The interesting jump is F3 to F4: from agents you delegate to, to a factory you direct.
Fig. 1 · Human share of the effort behind a merged changeSchematic, by autonomy level. Hover for values.
F3 · you are here
F0
Manual
Most of software history
Humans write every line. Tools compile, lint, and complain. The keyboard is the factory.
Gates
▸None. This is the floor.
F1
Assisted
2021 to 2023
Autocomplete gets good. A model suggests the next few tokens, but nothing ships without a human typing most of it.
Gates
▸Model-suggested code appears in a majority of new files.
▸Nobody reviews differently because of it.
F2
Supervised
2024
An agent writes whole functions and files while a human watches every step and approves every command. One agent, one human. Faster, but still serial.
Gates
▸An agent completes multi-file changes end to end.
▸The human reads every line before it merges.
▸One human drives at most one agent at a time.
F3
Delegated
2025 to nowWhere serious teams are
The agent owns a task from ticket to diff. The human reviews the result, not the keystrokes. Attention shifts from writing code to specifying work and judging it. One engineer runs several agents at once in isolated workspaces.
Fleets of agents plan, implement, review, and test each other's work. Humans set direction, arbitrate exceptions, and own taste. The unit of human attention is no longer the pull request. It is the decision.
Outcome in, software out. Humans specify intent, constraints, and budget. The factory schedules itself, ships continuously, monitors what it shipped, and rolls itself back. Most changes merge with no human in the loop, and the incident rate does not rise.
Gates
○A majority of production changes merge without any human reading the diff.
○Change-failure rate at or below the human-era baseline for two consecutive quarters.
○The factory reverts its own bad deploys faster than a human on-call did.
○Humans in the loop are there for judgment calls, not throughput.
The forecast
How 2026 plays out
Part record, part bet. The first two entries are already happening. The rest is stated concretely enough to be wrong about.
Fig. 2 · Merged changes written by agents, zero human editsIndustry median, our estimate. Dashed is forecast. The first F4 teams cross the gate earlier.
Early 2026
Early 2026
Review becomes the bottleneck
Happened
Writing code is no longer where engineer hours go. Reading it is. Teams that adopted parallel agents in 2025 hit the wall first: ten agents can produce more diffs before lunch than a team can honestly review by Friday.
The response: agents review agents, and humans sample instead of reading everything. Trust is earned statistically, not per diff.
What has to become true
Agent reviewers catch planted regressions at parity with median human reviewers in blind tests.
Mid 2026
Mid 2026
The dispatcher appears
Underway
The job description shifts. Engineers stop being typists with taste and become dispatchers with taste: decomposing work, routing it to fleets, arbitrating conflicts between agents that both touched the same module.
Tools that treat agents as a fleet, not a chat window, become the default interface to the codebase.
What has to become true
One engineer sustains 10 or more concurrent workstreams, and merge-conflict resolution between agent branches is itself mostly automated.
Late 2026
Late 2026
The overnight shift
Forecast
Long-horizon reliability crosses a threshold. Work assigned at 6pm is mergeable at 9am often enough that not scheduling the overnight shift feels like leaving a factory idle.
Environment setup, flaky tests, and credential plumbing, the boring failure modes that killed unattended runs in 2025, are mostly engineered away rather than modeled away.
What has to become true
Unattended runs of 8 hours or more succeed on a majority of routine tickets without a human unblocking them.
2027
2027
The first F4 teams
Forecast
The first teams, small ones, ship majority-agent code without reading every line, and their defect rates hold. Nothing mystical behind it: review layers, canaries, fast rollback, and a habit of writing specifications instead of code.
Everyone argues about whether this generalizes. That argument is the sign the level was reached.
What has to become true
At least one team we can name, ours included, passes every F4 gate for a full quarter and publishes the numbers.
Beyond: F5 is not a near-term claim
We do not forecast full self-driving in 2026 or 2027. The honest unknowns: whether agent review holds up against adversarial complexity, whether specification can replace code reading as the trust anchor at scale, and whether compute economics keep the overnight shift cheaper than the humans it augments. F5 is a rubric entry so we recognize it when we see it, not a promise.
The scorecard
Where the industry is, honestly
Our read as of August 2026, based on our own team and the teams we watch closely. F3 is mostly open. F4 is mostly closed. That gap is the work.
Standard for well-scoped work in isolated workspaces.
●Open
F3
3+ concurrent workstreams per engineer
Daily practice for our team and our heaviest users.
●Open
F3
Zero-edit majority on merged agent PRs
True for routine work, not yet for gnarly refactors.
◐Partial
F4
Half of merged changes are zero-edit agent code
Humans still edit or heavily steer most merged diffs.
○Closed
F4
Agent review at parity with human review
Agent review catches real bugs but is not yet trusted alone.
○Closed
F4
10+ concurrent workstreams without dropped state
Possible on good days. Supervision cost still grows too fast.
◐Partial
F4
Unattended overnight runs, mergeable by morning
Works when environments are clean. Environments are rarely clean.
◐Partial
F4
Ticket to production under one day, median
Review and CI queues eat the gains agents create.
○Closed
The factory needs a floor
Superset is the workbench for the F3 to F4 transition: parallel agents in isolated workspaces, fleets you can actually supervise, and a review surface for code you did not write.