Image: GitHub
Why it mattersMoving workflow orchestration from LLM inference to a deterministic script is where token cost actually comes down, and the Votito setup is one of the first public blueprints of what that looks like.
Pouring more tokens at a coding agent and hoping the output improves is where most teams start. Gojko Adzic, the agile and specification-by-example writer, has published the setup his own company Votito uses to go the other way. His new lights-on-software-factory repository pairs Claude Code with a bash regulator, a lint hook and a deterministic orchestrator, and he says the orchestrator cut Votito's token spend by around 80 per cent compared to running Claude Code with no guardrails.
The repository has gathered 88 stars since Adzic opened it this week. The pitch, from Adzic's README: "humans supervising agents, deterministic checks providing guardrails instead of markdown prayers, deterministic orchestrator scheduling tasks instead of LLM inference."
The four pieces
The setup bundles four things, each covered in its own doc. The container gives Claude Code a sandbox that shares only the workspace directory with the host. The bash regulator deep-inspects every bash command the agent runs, approves or denies automatically against project rules, and tells the agent what to do instead when it is blocked. The lint hook runs ESLint on every file Claude Code edits, with custom domain rules the project adds over time. The orchestrator turns a written plan into a Claude Workflow script with parallel workers and gates, so each step runs on a worker agent, a gate agent checks the merged result, and a check agent tries to refute what each step claimed.
The orchestrator is where the token saving lives. Adzic's framing in the README is that it "moves many common workflow orchestration tasks from inference (LLM) to deterministic tools (command-line) speeding them up and significantly reducing token spend." The 80 per cent figure is specific to Votito's bare-bones Claude workflow baseline, not a measurement any reader will see on their own project without running both.
What this is not
The 80 per cent is Adzic's own number, from one company's internal use, with no public before-and-after breakdown. He also says in the slides linked from the README that the repository is "both a simple boilerplate setup for new projects, and an educational demonstration of the core concepts of software factories," so a team adopting it starts with a trivial sample project and replaces its workflow with their own rules.
The container is Alpine with Node.js, make and git. The example workspace runs Jest and ESLint, and the agents are configured through .claude/settings.local.json. The licence is MIT.
What a team can take from it today
The deterministic parts around the agent are the piece most Claude Code setups are missing. A bash regulator cuts the stream of "approve this command?" prompts, which is where approval fatigue leaks token budget and reviewer attention in the first place. A lint hook turns a repeated review comment into a rule the agent gets on every edit, so the reviewer stops writing the same note. And an orchestrator written as a plan generator removes the LLM from scheduling, so the model only reasons about the parts that need reasoning.
Adzic ran the slides on these concepts at GOTO Copenhagen, linked from the README. The boilerplate is the first public implementation that goes with them, so a reader can match the slides against the actual code the setup runs.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
