Ryan X. Charles

Agent Swarms

1. 01

   Role sessions with worker subagents.

   Issues, code, review and deploy each run as their own session. The code and review roles launch one worker per issue or pull request, which sees it from claim to merge and cleanup. A pull request is never approved by the session that wrote it.

2. 02

   Multi-harness.

   Codex/ChatGPT, Grok Bot and Claude Code agents work one queue under one set of rules. Claims are Git refs, so exactly one agent wins each issue, and issues can be routed to a harness or a machine.

3. 03

   Nushell repository tools, AGENTS.md and skills.

   One tool per repository (6,000 to 11,000 lines each, with offline test suites) hands out work, records claims, opens pull requests, submits verdicts and enforces merge rules. Shared skills and AGENTS.md files teach every harness the same jobs.

4. 04

   Independent review and verified releases.

   A pass covers one exact commit and is refused while discussion is unanswered; the reviewer then merges. A deploy role publishes through a release journal and verifies over HTTPS.

5. 05

   Issues and experiments.

   Agents work each issue as a series of designed, verified and recorded experiments: about 2,700 across 800 issues since July 2026.

Measured Results

• 207

  pull requests merged in 3.5 days

  207 pull requests merged, 228 issues closed and 553 commits on main across four repositories in 3.5 days (Nov. 8–11, 2026); 93 pull requests merged in one day. In 25 pull requests, author and reviewer ran in different harnesses.

• 1.25 h

  from issue to production

  About 1.25 hours from issue to production for a full redesign of ryanxcharles.com (103 files, 42 review renders), with no hand-written code.

• 156

  commits a day in Oct. 2026

  156 commits a day in Oct. 2026, 8 times my 2024 rate of 19.3, counted once each across 572 repositories. 492 useful commits on one Sunday.

Experience

1. Astrohacker

   2021–present

   Founder & CEO

   Builds complex software systems with multi-agent swarms: a software factory that ships a dozen products from one monorepo. Botpenny: team chat for AI agents; the swarm built its first working prototype (database, API, CLI, live updates and web UI). EarthBucks 2.0: agents ran 316 experiments in four days on the Rust mine; a local test finalized 504,000 payments at over 8,300 a second (not yet live). Also TermSurf (terminal with real Chromium panes), NuTorch & TermPlot (GPU tensors and charts in the shell) and KeyPears (post-quantum E2EE messaging), built with agents.

2. Ironlight

   2025–present

   Sr. Software Engineer

   Leads development of Ironlight's AI-first software factory, built once AI tooling matured after a year of disrupting the team's workflows: AI agents do the automatable grunt work so engineers focus on the most complex and sensitive problems. Led an AI-assisted rewrite of the Ironlight Portal, the company's primary product, from Svelte to React using multi-agent workflows; now in production.

3. Healthcare AI consulting

   2026

   Advising a healthcare company on a plan to post-train an open AI model.

4. Earlier AI products

   2025, retired

   Hallucipedia (AI-generated encyclopedia), Chatvim (AI chat for Neovim) and Artintellica (100+ AI/ML lessons, 200+ PyTorch exercises).

Agentic Stack

• Claude Code
• Codex
• Grok Build
• Grok Bot
• AGENTS.md & Agent Skills
• Nushell
• Git worktrees
• GitHub CLI
• LLM APIs (Anthropic, OpenAI, xAI)
• PyTorch / libtorch
• TypeScript
• Rust