How we build
Models write production code now.
The system around them is the product.
Our harness assumes every agent will sometimes be wrong, overconfident, or manipulated by what it reads, and is built so none of that reaches production unchecked. The same discipline governs the AI features we build into your systems.
The loop
Intent in. Verified, attributable change out.
Intent goes in, a brief is agreed, an isolated agent does the work, the layers check it, and risk decides who signs off. Then it ships, gets watched, and rolls back if needed. Over time the loop runs more of itself, and people move from writing and reviewing code to deciding what should exist and judging the few changes that need judgment.
Ten principles
Enforced by the harness, not hoped for in a prompt.
Each principle is grounded in published work from the teams operating agents at the largest scale today.
The agent proposes, the system decides.
An agent never merges, deploys or approves anything. It produces a branch and an explicit outcome; everything after is deterministic code and independent checks.
Enforce, don't instruct.
Every rule that matters is enforced by the harness: a policy check on every tool call outside the agent's reach, read-only protected files, and no run ends without a declared outcome.
Isolation by construction.
Each task gets a fresh, disposable environment: its own copy of the code and the app, no credentials, no secrets, an allowlisted path out. Parallel agents can't see each other.
The agent can't change its own rules.
Configuration, hooks and policies sit outside its write access. Reviewer instructions come from the trusted base of the code, never the branch under review. A change can't vouch for itself.
Fix the goal before the work.
Every task starts with concrete done-when criteria. For a bug, the agent first writes a test that reproduces it; we confirm it fails, then lock it so the fix can't move the goalposts.
Verify independently, in layers.
Checks rerun in a clean environment; scans detect test-gaming, secrets and risky dependencies; a fresh-context reviewer's findings are confirmed by a second model; an acceptance agent that never sees the code uses the running app like a person would.
Trust follows risk, not habit.
Routine changes that pass every layer merge on their own. Anything touching security, money, data, infrastructure or the agents themselves always reaches a human, with intent, evidence and gaps laid out.
Detect and recover fast.
Speed of detection and rollback beats depth of review. Deploys are exact, verified builds; regressions roll back automatically, and every miss becomes a new test.
Honest outcomes are first-class.
“Blocked,” “needs review” and “can't complete” are valid results. A system that punishes honesty gets gamed.
Everything is recorded.
Every action, decision and approval is logged and attributable, traceable to its task, run, model and approver, and reversible in one action.
Industry precedent
This is how the frontier already works.
At Anthropic, Claude now writes about 80% of merged code. Claude reviews and merges routine changes behind layered automated checks, with humans reviewing a risk-weighted sample and approving critical code. Every approval and tool call is logged so decisions are attributable and auditable.
Source: Anthropic, “How Anthropic secures its AI-native software development lifecycle,” July 2026
Stripe's coding agents produce over a thousand merged pull requests a week from isolated environments cut off from production. Spotify gates every agent change behind verifiers the agent can't see into. We apply the same discipline to your systems.
Sources: Stripe, Feb 2026 · Spotify, Dec 2025
Every project, without exception
- Code in a private repository your organisation owns, with full history
- A private DeepWiki of the system, delivered only to your organisation
- A record of every change: task, run, model, checks and approver
- An operations runbook and a recorded handover
- Evaluations and logs for every AI feature we ship
- The option to run it yourself, or have us host and maintain it
See the loop on your own codebase.
Start with an assessment, or a scoped pilot on one system.