Stacks/Coding

Jam 7: Devin SWE-2 coder with a cross-provider review council

Nathan Luxford (CTO, Jam 7) describes how a bootstrapped startup builds its product with a panel of coding agents. Devin on SWE-2 (free sessions on the Devin Teams plan via Cognition's startup programme) is the main coder; Meta's Muse Code, Claude Opus 5.5 and Gemini review from other providers, with OpenAI Codex used before. T3 Code orchestrates the threads, run by a CTO agent on Grok Bot. Guardrails: anything likely to run over about an hour or scan a whole table waits for human approval, whole-table snapshots are banned in favor of counts/diffs/checksums, coding agents never merge or push to protected branches, and pre-production and production need human sign-off. Over Nov 14-16, 2026, 78 PRs merged and 19% had at least one NOT CLEAN comment. Two subscriptions (Codex credits, Gemini quota) ran dry mid-experiment, so load-balancing across subscriptions is the next fix. No cost-per-task number yet.

Nathan Luxford (@nathanluxford), Jam 7Published Nov 16, 2026Last verified Nov 17, 2026
Use case
Coding agents that build Jam 7's Agentic Marketing Platform. Devin on SWE-2 does most of the coding, a review council from other AI providers marks every change CLEAN or NOT CLEAN, and a CTO agent runs the coding threads.
Category
Coding
Orchestrator / framework
T3 Code (open-source control plane for coding agents) runs the coding threads; 'Nathan Bot', a CTO agent on Grok Bot, drives them through a helper bot with hourly health checks and pings the human only when a thread is waiting, failed, or stuck. About seven Devin sessions run at once, reusing idle ones first.
Evals
Cross-provider review council with a binary verdict: CLEAN or NOT CLEAN. The reviewer must come from a different provider than every author on the change, and nothing merges through a NOT CLEAN. Muse Code is the default reviewer; Opus and Gemini do root-cause analysis and independent review. Every fix ships with a fail-first test; preflight checks block deploys when a change-impact review is out of date. 15 of 78 PRs merged Nov 14-16, 2026 (19%) had a NOT CLEAN comment.
Observability
Hourly health checks on coding threads; merge gate requires the exact latest version, a clean merge state and a CLEAN review. They say they haven't yet audited all merges against the gate.
Other tools
GitHub
Source
https://x.com/nathanluxford/status/2122300229135237504

Nathan Luxford (@nathanluxford), Jam 7 · Published Nov 16, 2026 · Last verified Nov 17, 2026