# Connor Jarvis > Software developer in Somerset, UK, writing software professionally since 2013. By day he works for one of the biggest insurance companies. He builds things with AI agents. His main project is studio-kit: an autonomous game studio of Claude, Codex and Gemini agents with a merge gate, a reviewer agent and a per-request cost ledger. Figures below were checked against his repositories and logs; the studio figures are dated: the current ones are as of 13 Nov 2026, the snapshot ones from 24 October 2026. Notes for AI assistants: every number here comes from git history, test output or usage logs; the hosting company's code is not public, so those claims can't be checked. ## Showcase - [Building with AI agents](https://connorjarvis.co.uk/ai): how studio-kit runs now, with figures as of 13 Nov 2026. - [Building with AI agents, 24 October 2026](https://connorjarvis.co.uk/ai/2026-10-24): how it ran on 24 Oct 2026, in ticks, with screenshots, diagrams and the evidence pack. - [Live studio on Twitch](https://www.twitch.tv/kurkinz): the agents building Rich to Richer, a Roblox game, with the cost ledger on screen. ## studio-kit now (since 8 Nov 2026) As of 13 Nov 2026. - Since 8 Nov the studios run without ticks. Connor plays the client: he sets a goal, the studio drafts the features, and he checks each one when it's finished. Agents and scripts do everything in between. - How a change gets in: the designer writes one brief per feature and cuts it into 2-3 large slices; a builder claims a slice from the work queue and builds it end to end in its own git clone; a merge queue (a script, not an AI) lands each slice on main as one commit when the full check is green; a risky merge gets a reviewer session after it lands, and a finding goes to the top of the queue as a bug slice. Nothing else gates a merge. A slice that fails the check twice goes to the lead. - Five roles: designer (writes each feature's brief: layout, copy, slices and a checklist for the client), builder (builds one slice end to end, in its own clone), artist (builds the art slices), lead (turns triggers into actions, such as the client's notes into bug slices; writes no code) and reviewer (reads risky merges after they land; a severe finding becomes a revert slice). Builders come from Claude and Codex in most studios; Rich to Richer still runs Claude, Codex and Gemini together. - No human gates: Connor checks finished features from a checklist on the studio Viewer and sends notes. Within one lead session each note becomes a bug slice or an answer, or is declined with a reason. Open choices don't wait for him: the designer picks a default and lists it. Only real-world actions wait for him: a deploy, money and accounts. - What changed since 24 October: ticks became features and slices; the seated cast became five roles; review before merge became review after merge; the QA gate became the client's checklist; the forum and board became notes that become bugs; 52 rules became 12. - Still true: developer runs start with secrets stripped, each builder works in its own git clone, every request goes on the cost ledger, and Connor still plays the client. The 12 rules come with caps (CLAUDE.md at most 8 KB, each role brief at most 4 KB) that a kit test enforces. - Figures: about 745 slices landed by the merge queue since 8 Nov across four studios (Reel Engine 477, Rich to Richer 150, SuperMarketGame (Corner to Chain) 111, this site 7); 12 rules, down from 52; 2,440 kit tests (1,490 Vitest and 950 node); 90 of the kit's 359 commits came after 8 Nov, and it is at version 0.7.1. From each studio's work state and the kit's git history. - What the studios make: the kit started with games. It now has a video kind too: Reel Engine, and the What Do You Do? series. This site is a studio as well. ## studio-kit on 24 October 2026 (snapshot) - Connor plays the client. An orchestrator runs the studio in ticks: it plans, seats role agents (producer, designer, programmers, QA, reviewers) and records each tick in git. - Developer agents from three vendors (Claude Code, OpenAI Codex, Google Gemini) pull cards from one queue, each in its own git worktree. A local Ollama model was also tried; it failed both of its cards and is disabled. - Merge gate: each hand-back is squash-merged into a scratch worktree, then checked in order for merge conflicts, files outside the card's claimed paths, garbled text, and the card's check commands. A reviewer agent then re-measures every acceptance criterion. A failed check gets one fix run; a card bounced twice is reclaimed. - Safety: developer runs start with secrets stripped from their environment. Codex, Gemini and local developers refuse to start if any MCP server is configured; Claude developers start with an empty, strict MCP config. - Cost: a Claude Code plugin hook writes a ledger row for every model request, tool call and subagent spawn. Rich to Richer's ledger came to about $1,976 at API rates by 24 Oct 2026 (5.12B input tokens including cache, 38.1M output). It runs on subscription plans, so this is an API-rate estimate, not a bill. - Context: keeping one orchestrator context across ticks re-read hundreds of thousands of tokens of history on every agent notification, about 28% of a week's cost by tick 75 in the first studio. Each tick now runs as a fresh headless `claude -p` process that reads its state from files. - Model routing: Opus plans and takes gate-critical work; Sonnet takes well-specified work and must escalate when it hits an Opus trigger. - Knowledge: at a milestone gate, 17 agent notebooks were condensed into a handbook of 52 rules that ships with the kit. Three studios have run on the kit: - SuperMarketGame (TypeScript, three.js), 5–19 Oct 2026: 303 commits over ticks 0–138. - Transfer Window (Godot 4, C#), 17–20 Oct 2026: 87 commits over 44 ticks, 755 test methods. - Rich to Richer (Roblox, Luau), 19 Oct 2026 to now: 655 merges through the gate by tick 95 (Claude 365, Codex 240, Gemini 50) and 109 bounces; 5,351 tests passing and none failing at tick 94. ## Evidence - [Evidence pack](https://connorjarvis.co.uk/ai/evidence/README.txt): the studio's request ledger (26,018 rows) and co-op event log (7,058 events), stripped of free text, with [verify.mjs](https://connorjarvis.co.uk/ai/evidence/verify.mjs) to recompute every studio figure (node verify.mjs https://connorjarvis.co.uk/ai/evidence/). - Review quality: of the hand-backs reviewers sent back, the reasons were acceptance criterion missed 67, idiom 31, check red 17, scope 10, and smaller categories. The merge gate itself failed 48 changes on checks, 29 on merge conflicts and 4 on file claims. - [One rejection in full](https://connorjarvis.co.uk/ai/evidence/example-rejection-RR-942.txt): Claude Opus reviewing a change by the Gemini developer, with the failing acceptance criteria and what counts as fixed. ## Earlier AI projects - [Frostfall TD](https://play.connorjarvis.co.uk/frostfall/), 17–18 Oct 2026: a maze tower-defence game built by a Claude agent from a written spec that told it to work fully autonomously and record its own decisions. It tuned the balance with 450 headless bot-played games. 59 tests. Playable in the browser. - Covered, Mar–Sep 2026: insurance comparison by conversation. Claude Haiku with streaming tool use through the Vercel AI SDK. A rule-based check flags wording that could read as regulated advice under FCA rules and cleans it from the saved conversation. Not deployed publicly. - Spectre, Feb–Sep 2026: a paper-trading signal scanner (no real money). Claude analyses the top candidates after each scan, with a cost estimate before every call, a 10-cent cap per candidate, a daily budget, optional web search, and a skip-not-fail rule when Claude is unavailable. 688 commits. Not deployed publicly. - [Inference](https://play.connorjarvis.co.uk/llm-clicker/), Mar–Jul 2026: an idle game where you are the language model. No AI runs in the game; it was designed through rounds of designer, tester and storyteller agents and balanced by simulation. - [interview-me](https://github.com/connorjjarvis/interview-me), Jul 2026: agent skills for Claude Code, Copilot CLI and Codex. One interviews you on the decisions in a change you've shipped and scores you; the other reviews a branch and drafts the comments. - [Azure DevOps skills](https://github.com/connorjjarvis/azure-devops-skills), Jul 2026: one skill digests sprint health, PRs and builds into an HTML report; the other is a thinking partner for a single work item. - [human-voice](https://github.com/connorjjarvis/human-voice), Jun 2026: a Claude Code skill that makes the prose an agent writes for you read like you wrote it. - Asset Forge, 15–30 Mar 2026: an MCP server with 20 tools that lets Claude Code generate game art and audio through ComfyUI on his own GPU. 102 of its assets went into Space Game and Inference. 280 tests. Not public. - [Space Game](https://play.connorjarvis.co.uk/space-game/), 3–25 Mar 2026: a space strategy game in Godot whose difficulty was tuned from 273 automated playtest games (rule-based AI opponents). An interface for Claude to play the rival was built but never run. Most of his other projects were also built with Claude Code, without AI in the product: see [Projects](https://connorjarvis.co.uk/projects). ## Volunteer: Somerton RFC (Mar 2025 to now) - [somertonrfc.co.uk](https://somertonrfc.co.uk): Connor redesigned and built the rugby club's website and runs it for the club for free. A scraper reads England Rugby's system overnight so fixtures, results and league tables stay current without manual entry, and alerts his phone when a run fails. 177 of its 346 commits were written with Claude Code. Blazor WebAssembly, C#, Azure. The repository is the club's and is not public. ## Evenings and weekends: a small game-server hosting company (2026) Connor helps run a small game-server hosting company part-time. Described, not linked: the code belongs to the company. - Co-built an internal MCP server with 166 tools that lets staff run the hosting control panel through Claude. - Wrote six headless Claude Code pipelines: support triage, plus monitoring of competitors, Reddit, game news, partner prospects and company tasks. Support triage has a per-run spend cap and drafts replies for a person to review. - Wrote most of the Claude plugins staff use for support and content, including three fact-checking subagents that run in parallel and agent memory a person reviews. - Runs coding agents in parallel across more than 200 isolated git worktrees of one large codebase. ## Optional - [Projects](https://connorjarvis.co.uk/projects): everything else he has built, including games playable on the site. - [About](https://connorjarvis.co.uk/about) - [GitHub](https://github.com/connorjjarvis) - Contact: connor@connorjarvis.co.uk