Manifesto
The bottleneck moved. Most people haven’t noticed.
For fifty years the scarce thing in software was a person who could write it. Whole industries organized around that scarcity: the hiring funnel, the sprint ritual, the seniority ladder. That constraint is gone. AI writes three quarters of the new code at Google now, and it writes it well.
We felt it ourselves. The first time an agent implemented a change faster and cleaner than we would have, the honest reaction wasn’t fear. It was relief, followed by a question nobody had a good answer to. Fine, it wrote the code. Who answers for it?
The bottleneck didn’t disappear. It moved from writing the code to being accountable for it. Most of the industry is still optimizing the part that’s already solved.
AI share of new code: Google, Q3 2025 earnings call
02Ninety five percent
If code were still the constraint, the numbers would show it: everything ten times faster, ten times cheaper. Instead, ninety five percent of enterprise AI pilots show no measurable P&L impact. The models can code. The pilots still die.
They die at a predictable spot. The demo ships beautifully on a clean repo. Then someone in the real organization asks the only question that matters: who approved this? There is no answer. No owner, no chain of approval, no record that survives an audit.
The bottleneck isn't writing code anymore. It's accountability.
03The babysitting moved up a level
A copilot makes one engineer faster. A task agent takes a ticket and returns a pull request. Both are real progress, and both share a failure mode: every line they produce still funnels through one human who must read it, understand it, and stake their name on it.
Scale the agents and you scale the queue, not the throughput. Ten agents filing pull requests at one reviewer is not a faster team. It is a slower human. The babysitting didn’t go away. It moved up a level, and it concentrated.
04Accountability is a structure
Accountability is not a feeling of trust in the model. It is a structure you can point at. An identity, so every action has an owner. Gates, so nothing ships unproven. A tamper evident record, so the history is either intact or provably altered. A boundary, so a human decides exactly the things a human must decide, with a kill switch one click away.
Anthropic sells the engineer's brain. We sell the organization the brain works inside.
05Fourteen gates. Three are human.
Every change tOOrunt AI ships crosses the same fourteen gates, in order, no exceptions. Eleven are machine cleared: right repository, requirements understood, tests green, zero quality findings, no secrets in the code, risk scored against a cap. Three are human: the plan is approved before any code is written, the pull request is signed by a reviewer, and the merge happens only on verified green.
That is the entire human job. Three decisions per change, placed exactly where judgment matters and nowhere else.
- 01Right repoautoresolves the target repo from the ticket, using a curated registry or semantic match
- 02Requirementsautono material ambiguity. Clarifies with a human before building and never guesses
- 03Dependenciesautonothing this change depends on is missing; parks and resumes if it is
- 04Plan approvedhumanthe plan is posted to Jira and waits. It never writes code before a human sees it
- 05Noveltyautoroutine work ships; unfamiliar territory forces extra deliberation and a human merge
- 06Reworkautowhen a gate pushes back, it improves the change and runs again. It does not give up
- 07Tests greenautothe change's own suite passes; for a bug, a red to green reproduction proves the fix
- 08Code qualityauto0 findings: no leaks, deadlocks, bug patterns, or injections
- 09Coverageautoeach changed file has covering tests; missing ones are written before the PR
- 10No secretsautoscanned for live credential values and secret patterns: clean
- 11Risk capautoscores risk and confidence; anything past the cap escalates to a human
- 12Review signedhumana second agent reviews a 7 point checklist; a human reviewer signs the PR
- 13Mergedhumanmerges only on verified green tests, CI, no conflicts, and review approval
- 14Watchedautowatches CI and production after merge; raises a revert alert if it regresses
06A record you can’t fake
Each audit record commits to the hash of the one before it. Alter any past decision and every subsequent hash breaks visibly and provably. There is no third state between intact and tampered. Incident forensics and SOC 2 evidence stop being two documents; they are the same artifact, HMAC signed and exportable.
We hold ourselves to it. The founding run in the next section lives on that chain. We could not have faked it after the fact if we had wanted to.
- 03:05:12toorunt sam plan_approved#9f3a1c ← 0ac913
- 03:05:04founder approve_prd#1b77e0 ← 9f3a1c
- 03:04:51toorunt sam tests_green#c40d92 ← 1b77e0
- 03:04:39toorunt tyler peer_review:approve#77a5f1 ← c40d92
- 03:04:22toorunt sam secret_scan:clean#2e9b04 ← 77a5f1
- 03:04:08toorunt sam merge:verified green#b81c6d ← 2e9b04
- 03:03:55toorunt sam watchdog:armed#5d2af8 ← b81c6d
- 03:03:40founder kill_switch:off#0ac913 ← 5d2af8
- 03:05:12toorunt sam plan_approved#9f3a1c ← 0ac913
- 03:05:04founder approve_prd#1b77e0 ← 9f3a1c
- 03:04:51toorunt sam tests_green#c40d92 ← 1b77e0
- 03:04:39toorunt tyler peer_review:approve#77a5f1 ← c40d92
- 03:04:22toorunt sam secret_scan:clean#2e9b04 ← 77a5f1
- 03:04:08toorunt sam merge:verified green#b81c6d ← 2e9b04
- 03:03:55toorunt sam watchdog:armed#5d2af8 ← b81c6d
- 03:03:40founder kill_switch:off#0ac913 ← 5d2af8
07Two hours, thirty six minutes
One evening in July, a one line product idea went in. Two hours and thirty six minutes later, a product was deployed: a signed spec, a backlog, thirty plus rounds of review between two agents with distinct GitHub identities and exactly three human decisions along the way. Every step is on the hash chained record.
- idea to deployed product
- 2h 36m
- human decisions
- 3
- gates per change
- 14
- compute per merged PR
- $20 to 150
Not a benchmark. A receipt.
From idea to shipped software
04Idea → shipped software · humans decide three things
- IdeaThe same five phase pipeline runs whether you hand tOOrunt AI a one line product idea or a ticket off your existing backlog.
- HUMANPRD signedIt posts an implementation plan to Jira: files, approach, risks, and acceptance criteria, then stops. It never writes code before a human sees the plan.
- BacklogIt watches your board, picks up the ticket, and resolves the right GitHub repo from a curated registry or a semantic match. Ambiguous requirements get a clarifying question, never a guess.
- Build + testsIt implements the change and writes tests until green. For a bug, it writes a reproduction test that is RED before the fix and GREEN after. This proves the fix actually fixes the bug.
- Peer reviewA second agent reviews a seven point checklist; a peer bot with a distinct identity reviews the PR. Human comments are handled one by one: fixed, politely rebutted, or asked about.
- HUMANPR approveA peer bot with its own GitHub identity reviews the change in a real, adversarial pass, then waits for your approval before the second gate opens.
- HUMANMerge + watchIt merges only on verified green tests, CI, no conflicts, and review approval, then watches CI and production, raising a revert alert if the change regresses.
3 human decisions: sign the PRD, approve the PR, unlock the merge · 14 hard gates on every change · everything between them runs autonomous, on the record
08The economics, in monospace
A merged pull request from a senior engineer costs the organization five hundred to a thousand dollars once salary, review time, and coordination are counted. The same unit of work through tOOrunt AI is twenty to a hundred and fifty dollars of metered compute on the ledger, per PR, visible. The price is a hundred and fifty dollars flat, per gate passed, complexity normalized merged pull request.
The unit is the point. Not seats, not usage vibes: a priced, audited unit of shipped work. Engineering stops being a cost you estimate and becomes one you read.
09The shape of the team after
The backlog was never too much work. It was work waiting behind too few people entitled to say yes. When saying yes is structured, with three decisions, fourteen gates, and a chain of custody, the queue drains, and the org chart changes shape.
The pyramid of implementers flattens into a wide, governed, autonomous interior. What remains of the human organization is thin, senior, and load bearing: the people who decide what should exist, and sign for it.
Somewhere tonight a backlog is draining while its owners sleep. Every change owned. Every gate on the record. Three decisions that were actually theirs.
Worst case is a rejected pull request.
The tOOrunt AI team · July 2026, on the chain