Ratchet Agentic Engineering. The process that scales AI agents without losing quality.
RiftSeed is a software engineering lab. It develops a method for building software with AI agents under control, the tool that applies it, and tailored platforms built with the method.
The method’s rules leave quality one direction only: up.
90% of software development professionals use AI at work; 30% have little or no trust in the code it generates. 1
How a change goes round
Four actors: a person, Cockpit, the code agent and CI. Every change passes through all of them, in the same eight steps.
Request (Person → Code agent → Cockpit): The request is logged in the words of whoever made it and checked against the requirements that already exist.
Requirement (Cockpit → Person → Cockpit): Agents propose requirements; only a person confirms them, in Cockpit.
Impact (Cockpit → Code agent): Before touching the code the agent asks for the impact analysis: files, modules and tests to run, each with its reason.
Agents (Code agent): Every session starts from a briefing (priorities, open requirements, failing tests) and ends with commits, files and tests on record.
Gate (Code agent → CI and production → Cockpit): The full test run on every push; the ratchet ceilings are held by the server and can only go down.
Person (Cockpit → Person): The verdict on the release comes from a person, in Cockpit or in the session: no agent approves.
Release (Person → Cockpit → CI and production): Every release stays on record with its version, its verdict and its notes.
Probes (CI and production → Cockpit → Code agent): A failure becomes an incident, which closes only with the guard test that reproduces it.
Red or rejected: back to the agents. Annotated lesson: a new test and a dated note, into the next round.
A fault leaves a test behind, and the ceiling never goes back.
A fault in production: After midnight the journal’s week comes up empty.
The probe sees it: The probe replays the scene on the live site and the fault becomes an incident.
A guard test reproduces it: Before the fix there is a test that fails for the same reason.
The ceiling tightens: The incident closes only with that test. The ceiling drops and does not rise again.
The place where code agents receive the work and leave the evidence: requirements, impact analysis, tests, verdicts. A person confirms the requirements and decides every release. Prototype in development, private preview.
The method: Ratchet Agentic Engineering. The rules under which code agents work and account for their work. Today: In use on every project of the lab. The public specification is at version 0.9, in draft. (In use)
The tool: Cockpit by RiftSeed. Where agents receive the work and leave the evidence, and where a person decides. Today: A prototype used every day on the lab’s own projects, in private preview. (Prototype in development)
The platforms: Tailored platforms. Cloud software with AI inside, shaped on the work of the people who use it and shared across organisations with separate data. Today: Built and maintained with the method, on the lab’s shared base. (In use)
Now
The method in use on every project of the lab.
Cockpit used every day in the lab.
Next
Specification 1.0, with its levels.
Cockpit for more people.
Later
In Cockpit, the developer puts together their own agent team, with Claude Code and Codex side by side.
Tailored platforms for organisations with processes of their own, built with the method.