🤝 Top Collaborators
🤓 Latest Submissions
Tablekeeper: AI-Built Restaurant Reservations
Tablekeeper is a restaurant reservation system with one hard rule: a table is never double-booked. It was built entirely by a "dark factory" of three AI coding agents sharing one BAND room, using Claude Code with the claude-sonnet-5 model. No human wrote any application code. The Planner reads the task and spec, writes a requirements checklist and assigns work items. The Implementer writes the code and tests, builds the Docker image, runs the official harness and posts evidence. The Reviewer writes no application code: it rebuilds the container and re-runs every check itself, attacking the service with replayed idempotency keys, double-booking attempts and daylight-saving changes, then accepts or rejects. Stage 1 is a JSON API for restaurants, tables, availability and reservations. Stage 2 adds a browser website (search by restaurant, date and party size, book, look up by reference code, cancel) and combinable tables for bigger parties. Both stages pass the official harness in isolated mode (stage 1: 120/120, stage 2: 25/25 shipped checks), our own suite of 282 tests passes, and each image starts in a fresh container with no network. The repository includes the agent mandates, FACTORY.md describing how the factory works and catches bad work, and the full BAND room export. We are honest about limits: no per-seat git history, no visual inspection of the site, and stages 3 and 4 were not attempted.
6 Oct 2026
ReviewSquad: AI code review with IBM Bob
The problem Code review is one of the slowest and least reliable steps in software delivery. Pull requests wait hours for a busy reviewer, and when the review happens, it is often rushed: the tests pass, the diff looks reasonable, approved. But passing tests do not mean safe code. Hard-coded secrets, SQL injection, inverted business logic and untested code slip into production, where fixing them costs far more. Reviews are also inconsistent: each reviewer checks different things, and the team's written guidelines are rarely applied line by line. The solution ReviewSquad turns IBM Bob into a disciplined code review team. A developer opens the pull request in IBM Bob, selects the Review Lead mode and asks it to review. ReviewSquad then runs the full workflow: 1. It converts the git diff into a review brief with exact line numbers and runs the tests and fast automated checks for hard evidence. 2. The Review Lead reads the team's own code review guidelines from a PDF (document understanding). 3. It runs four specialist reviews - Security, Bugs & Logic, Tests and Team Rules - each recording structured findings with the file, line, severity, the guideline rule broken, and a fix. 4. The findings are merged into one review with a clear verdict. 5. After a human approves, a Fixer mode repairs every finding and adds tests; tests and checks run again. 6. ReviewSquad publishes a shareable report of what was found, by whom, and the before/after state. Target users are development teams where senior reviewers are a bottleneck. Developers interact entirely through IBM Bob's chat; reviewers read one report. Why it is different Most AI review tools make a single pass and leave a list of comments. ReviewSquad runs the whole workflow - understand the rules, review, decide, fix, verify, report - with clear roles and permissions: reviewer modes are read-only, and only the Fixer may edit code, only after approval. Deterministic tools do the counting; Bob does the judgment. Because the rules come from a document, every team gets reviews against its own standards, not generic advice. We also measure instead of claiming. Our demo pull request contains 11 planted problems and passes all 8 of its tests; GitHub marks it "Ready to merge". A hidden answer key, which Bob cannot read, scores the review. Impact (our recorded run) - Automated checks alone caught 7 of 11 planted problems; ReviewSquad with IBM Bob caught 10 of 11, including an inverted coupon-expiry check that no test or linter flagged. - After the Fixer: 37 of 37 tests pass (up from 8), and automated warnings dropped from 22 to 0. - The full review, fix and verification took minutes of agent time instead of a manual review cycle, with every finding traceable to a line and a team rule.
27 Sep 2026
