◇ agi-cli · field notes
Blog
Running coding agents on machines you control. Case studies, comparisons, benchmarks, and the parts of the setup that aren't obvious.
Know what your agents spend
The balanced strategy sprays every run across every account and harness you have — which is also why the bill is scattered. agents cost sums it, agents insights splits it by account, and agents budget caps a run before it starts and kills a headless one that crosses the line.
One engineer, a factory of agents
Taylor Gagne is the principal engineer at YC-backed Turing Labs. He built Cybertron — a spec-driven orchestrator that routes each task to the best model and harness — and runs it on agi-cli across four harnesses and every machine he owns. Then he sent four patches back upstream.
An agent bought its own code-signing certificate to unblock Windows
Closing the loop: Windows SmartScreen was killing installs, so the agent bought a code-signing cert on its own. It compared EV vs OV, made the account, paid $129 from 1Password without ever printing the card, and filed the whole order. No human in the loop, not even for the purchase.
An agent designed a part and 3D-printed it — from one spoken prompt
One voice brief for a mic clip. The agent modeled it, pulled the real headphone mesh to check the fit, sliced it, caught its own slicer warning, checked the printer's camera, and sent the job.
Version-pin your coding agents like .nvmrc
agents.yaml at the repo root pins each agent CLI's exact version. Every agents call inside the tree honors it, and agents check fails CI when a machine drifts.
Dodge rate limits by rotating across accounts
The balanced strategy spreads each run across the signed-in accounts with capacity to spare and skips any whose rate-limit window is maxed out.
Search every agent transcript from one place
Every run writes a transcript to one indexed store. Search it across Claude, Codex, Gemini, and OpenCode, replay it, and query other machines over SSH.
Sync AGENTS.md, CLAUDE.md and MCP config across every AI agent
The same rules end up copy-pasted across five files and drift within a week. Keep one source of truth and let agi-cli write each agent's file.
A source-available, self-hosted alternative to Factory.ai
Same category, source-available counter-position: agent-agnostic, self-hostable, and your code never leaves your infrastructure.
agi-cli vs Cursor
An AI-native editor vs a source-available, self-hosted CLI that runs any coding agent (Claude Code, Codex, Gemini) on the subscription you already pay for.
agi-cli vs OpenCode
OpenCode is one agent. agi-cli is the harness that runs it and every other agent CLI, on machines you control.
agi-cli vs running CLIs by hand
The same native CLIs you already run, plus version pins, account rotation, session search, and Keychain-backed secrets.