Krizaka Agents: how we ship with coding agents
For peers — What surrounds Claude Code to ship work: one session per ticket, checks that refuse, a pull request, a person who decides.
For everyone — How we build our products: AI agents write the changes, automatic checks verify them, a person accepts or refuses them.
We build our products with it, every day. It is not for sale for now.
Status: in development · internal use. Works with Claude Code.
§ 1 — What it is
Krizaka Agents is the tooling around the models: a launcher that starts every session the same way, guards that refuse, trains that merge, a watchdog that reads, a cockpit that shows. The sessions run in Claude Code; Claude Code provides the agent loop, not us. What we build is what surrounds it: what delivers each change as a pull request and has it refused if it does not pass.
All of it serves one job: delivery. It starts at the ticket and ends at the merge and its verification.
§ 2 — Three rules, and the rest follows
- GitHub is the source. What we keep locally is a copy, to show and to warn.
- No model where the answer must be certain. The launcher, the guards, the trains and the watchdog are plain code.
- Agents do not talk to each other. A session writes what it has to say somewhere that stays: a commit, a pull request, a row in a database.
§ 3 — Who decides
A person sets the direction and decides what is theirs: money, accounts, legal matters, what cannot be undone. A CTO agent decides the rest as long as it can be reversed, with a written cap whenever a spend is possible. Agents propose every change as a pull request; they do not merge.
§ 4 — Where we are
It is our own technology, the one we build our products with, Sovkora first. It changes every week. It is not on sale and not self-service, and its code stays private. This site shows the method; there is nothing to buy or to request here.
Not proven yet
- The link between overlapping work and rework, in time or money.
- Any of this with an agent other than Claude Code.