slabgorb.com
the home page of keith avery :: engineer & artist
Pennyfarthing
// Claude Code agent-orchestration framework — v13+
A framework that turns a team of specialized AI agents into a single, coordinated engineering crew. I built it at 1898 & Co., then used it to build everything else I shipped there.
One engineer prompting one AI assistant hits a ceiling fast. The context falls out of the window, behavior drifts between sessions, and there's no way for the work of one task to inform the next. For a critical-infrastructure consultancy with real engineering teams, that's not good enough.
I wanted something different: a crew of specialized agents that hold consistent personalities, hand work between each other along defined routes, and remember what they've done — so the whole system gets more capable over a sprint, not less.
Pennyfarthing is built on Claude Code and MCP. Four moving parts do the heavy lifting:
The framework's first real client was itself. By dogfooding Pennyfarthing to build Pennyfarthing, the system delivered 4,468 commits in 12 weeks — about 372 a week, or roughly 20× typical development velocity.
To make sure that velocity wasn't noise, I benchmarked agent behavior scientifically: OCEAN personality profiling, the TRAIL evaluation framework, and Cohen's d effect sizes measured across 45 persona themes.
Claude CodeMCPPythonGolangFastAPIuvicornBuilt for an infrastructure client — the codebase is internal. Happy to walk through the architecture in detail.