Feed it, staff it, sell it, sheet 13 of 13
The loop around the model: tools, permissions, context, memory, sub-agents and scheduled runs, the choice between adopting a coding agent and building your own, and running several at once.
Keep your house rules in AGENTS.md and your permissions in one written policy, so the harness you pick today is easy to change next quarter.
← → move to the next sheet over, Esc zooms out.
A coding agent is two things: the model that does the thinking, and the harness wrapped around it. The harness is everything else. It runs the loop that lets the model call a tool and read the result, it holds the permissions, it decides what to summarize when the context fills up, it starts helpers, and it can run the whole thing on a schedule with nobody watching.
My day job is me and a staff of agents. The model is who I hired. The harness is the office I set up around them: which keys they get, the notes on the wall they read every morning, who they can hand work to, and the checklist the night shift follows. Same hire, different office, different week.
When you pick a coding agent you're mostly picking a harness, because the same model can sit inside several of them. And when you build an agent into your own product or operations, the harness is the part you're building.
Then you run more than one. Each agent gets its own copy of the code on its own branch, and the job changes from writing to supervising: which agent is working, which is done, and which has been sitting there waiting on you. A new set of apps does the supervising, and they drive the same coding agents, Claude Code and Codex among them.
Every link opened on 19 Nov 2026. The numbers match the squares on the map.
Product, Anthropic
Anthropic's coding agent for the terminal, IDE and desktop, with hooks, mods, sub-agents that each get their own context window, skills, and project memory from CLAUDE.md or AGENTS.md.
The catch. Not open source. Its licence reserves all rights, and use falls under Anthropic's commercial terms.
Open source, Apache-2.0
OpenAI's lightweight coding agent for the terminal, which reads AGENTS.md for project instructions.
The catch. Its sandbox is on in the default mode. Know what full access turns off before anyone switches to it.
Open source, Apache-2.0
Google's open-source agent that brings Gemini into the terminal, with MCP support.
The catch. Open source means you can read what it does with your files. Somebody still has to.
Product, GitHub
GitHub's agents: a cloud agent that takes a task from a question to planned code changes and a review, and Copilot CLI in the terminal, with custom agents, skills and hooks.
The catch. Closest to your repos and your pull requests, and so closest to one vendor. Check what leaves with you.
Product, Cursor
Cursor's agent for autonomous coding tasks, terminal commands and editing, in the editor, the CLI and as cloud agents, with rules, skills, sub-agents, hooks and MCP.
The catch. Its rules and settings are Cursor's own. Keep the house rules in AGENTS.md so they outlive the editor.
Open source, MIT
An open-source coding agent for the terminal.
The catch. The project has moved GitHub organizations. Install from the current one, and pin the version.
Open source, Apache-2.0
An open-source, extensible agent that installs, runs, edits and tests code with any LLM. Now an Agentic AI Foundation project at the Linux Foundation, alongside MCP and AGENTS.md.
The catch. Any model means any model's habits. Your evals have to cover whichever one you plug in.
Open source, Apache-2.0
An autonomous coding agent you can run as an IDE extension, a CLI or an SDK.
The catch. Three ways to run it means three places to set the same permissions.
Open source, MIT
A minimal agent harness for the terminal that skips sub-agents and plan mode on purpose. You add what you need as extensions, skills and packages. omp is a fork of Pi with much more built in.
The catch. Minimal means the missing parts are yours to build, and then yours to maintain.
SDK, Anthropic
Claude Code as a library, in Python and TypeScript: the same agent loop for agents you build yourself, with cost tracking and OpenTelemetry observability.
The catch. Build on it and the harness is yours to host, secure and keep current.
Open source, MIT
A lightweight Python framework for multi-agent workflows.
The catch. A framework, not a coding agent. You bring the tools, the guardrails and the hosting.
Open source, Apache-2.0
An open-source, code-first Python toolkit for building, evaluating and deploying agents.
The catch. A kit for building agents, not a coding agent. The harness you build with it is yours to run.
Framework, Microsoft
A framework for building agents and orchestrating multi-agent workflows, in .NET, Python and Go.
The catch. Strongest inside Microsoft's stack. Outside it, compare it with the lighter options.
Open source, Apache-2.0
An open-source SDK for building an agent harness and controlling it end to end, in Python and TypeScript.
The catch. Its repository was renamed to harness-sdk. Install from the current name.
Open source, MIT
A low-level orchestration framework for building, managing and deploying long-running, stateful agents.
The catch. Low-level by design. You get control, and you write more of the harness yourself.
Reference, Anthropic
How an initializer agent and an incremental coding agent keep making steady progress across many context windows, built on the Claude Agent SDK.
The catch. A model vendor writing about its own SDK. The pattern works in any harness.
Open source, MIT, paid plans
A Mac app and CLI that runs Claude Code, Codex and other coding agents side by side, each in its own git worktree and branch, with diff review in one window and an iPhone and iPad app for checking in away from the desk.
The catch. Renamed from Superset, so older guides use the old name. The phone app comes with the paid plans.
Open source, MIT
A desktop app for Mac, Windows and Linux that runs Claude Code, Codex and other agent CLIs in parallel, each in its own worktree, with diff comments you send back to the agent and a phone app that pings you when an agent needs you.
The catch. It can run agents on remote servers, SSH hosts and cloud VMs. Decide where your code may go before anyone adds a host.
Open source, Apache-2.0
A desktop and phone app, with a separate terminal version, for coding agents on your own machines and any you reach over SSH, where each agent reports whether it's working, done or waiting on you.
The catch. The status comes from an open protocol, OSC 7501. An agent that doesn't speak it still runs, just without a status.
Open source, Apache-2.0
A daemon on your machine that runs Claude Code, Codex, OpenCode, Pi and other agents, each in its own worktree if you want, with desktop, phone, web and command-line clients.
The catch. One person maintains it full time, by its own account, with no company behind it. Weigh that before a team depends on it.
Open source, Apache-2.0
OpenAI's spec, with an experimental Elixir reference build, for a service that reads work from an issue tracker and starts an isolated agent run for each issue. A run can end at a handoff such as human review. In OpenAI's demo the agents come back with proof of work and land the pull request once it's accepted.
The catch. OpenAI calls it a low-key engineering preview for testing in trusted environments. Its first suggestion is to have your own coding agent build it from the spec.
Reference, Stripe
Unattended one-shot agents, most often started from a Slack thread, each in a devbox isolated from production and the internet, with lints on every push and two rounds of CI at most. By Stripe's own count, over a thousand merged pull requests a week with no human-written code.
The catch. A person reviews every one. The factory ends at the door before the merge.
Reference, StrongDM
The boldest rules I've read: code must not be written by humans, and must not be reviewed by humans. Scenarios, often kept outside the codebase, and behavioral clones of the services it calls do the judging instead.
The catch. One team's rule, set by the three people who started it in July 2025, on a page that also asks for at least $1,000 a day in tokens per engineer. StrongDM is now part of Delinea.