How to set up an AI software factory for an engineering team
An AI software factory is agents that plan, code, test and review on your repos while engineers approve. How to set one up on your own models.
An AI software factory is a setup where coding agents plan, write, test and review changes on your repositories while engineers set the goals and decide what lands; CodeAF, the open-source (Apache 2.0) coding agent by AgentField, builds one in your terminal on the models you choose. This guide walks from an empty machine to teams with managers, specialists, house rules, budgets and nightly routines, in the order that keeps the work safe to hand off.
At a glance
| Factory piece | What it does | In CodeAF |
|---|---|---|
| Intake | Turns a sentence into a job that runs on its own | /task <brief>, in its own copy of the folder |
| Crew | Plans, does and checks each job | worker, planner and checker seats, each on its own model |
| Teams and managers | Groups agents around one outcome and hands out work | /teams (Alt+2), + Manager |
| Specialists | Built for one job each | /senior-dev, /sec, /review |
| House rules | Conditions every agent reads | /standing <words>, plus your CLAUDE.md and AGENTS.md |
| Routines | Work on a clock or a condition | schedules and watches, checked every 5 minutes |
| Money | Limits at every level | $500 a day backstop, $5 a task, team caps |
| Landing | Where finished work goes | merged into your folder, or kept on its own branch when you are on main |
Step 1: install and connect a model
curl -fsSL https://agentfield.ai/get/codeaf | bash
codeafWith no provider connected, the first run starts with choose a model provider, then opens Basic settings for your daily limit and chat model (first team). CodeAF itself is free; you pay your model provider (FAQ). The built-in specialists are left out of Windows builds.
Step 2: write the house rules once
/standing always write Go doc comments in full sentences puts up a card that reads wants to keep a rule; press 1 and every conversation in that project reads it on every turn, and applicable rules enter each task's brief as conditions (rules). CodeAF also loads a project's CLAUDE.md and AGENTS.md, and reads skills where they already live, so conventions you wrote for other agents carry over.
Step 3: set the money before the work
/budget 50 sets the day's limit; the default is $500 as a backstop (budgets). Each task is held to $5 unless you change it in /crew, and a team can carry a daily cap that covers it and every team under it. When a team hits its cap you are asked once, with Raise to $10 or Stop for today. A manager can never raise a cap.
Step 4: hand off the first jobs
/task write a pytest test file for each of cart.py, pricing.py and orders.py, one part per module starts at once while your chat stays free (fire and steer). The run works in a copy of your folder, splits into parts that run side by side, and a separate check reads each part against its acceptance. At the end the work is merged into your folder, or kept on its own branch when you are on a branch such as main that tasks never merge into.
Step 5: give the work a team and a manager
On Teams, click + New team, add conversations with + Add chat, then + Manager (first team). A manager is an agent whose job is the other agents: it hands out work, tracks each member, settles what it can and asks you only what it cannot decide (managers). Managers can run sub-teams, and a global manager can sit over every team. In our build a software company run, one message to a global manager over product, backend and qa teams produced a spec, the code and an end-to-end test in twelve minutes for about $0.70, with 14 tests passing.
Step 6: route the big jobs to specialists
/senior-dev takes one large, well-specified change and works alone on a branch of its own, then runs your project's build and tests on what it handed in (default ceiling $10 and 3 hours). /sec audits local code and /review owner/repo#123 reviews a GitHub pull request; both read code, run no tests and post nothing unless you ask (specialists). No specialist guarantees it found every bug.
Step 7: add routines that run without you
Say "every night at 2am, run go test ./... here, fix the easy failures and write what you did to MORNING.md" and a card offers to set up the schedule. A watch wakes on a condition instead, such as a failing test on main. Unattended runs work only under approvals you already gave (running unattended). For work that should outlast a laptop lid, see a long-running coding agent.
Step 8: decide how much it may do without asking
New conversations start in YOLO unless you saved another choice; Alt+A walks asks, guardian and YOLO, and dangerous commands and anything sent in your name always ask (FAQ). See approvals for regulated teams.
What is personal and what is shared
CodeAF runs on a machine: its teams, managers and routines live in the conversations on that machine, and home shows every project on it. Several engineers can reach one shared dev box over ssh with codeaf chat --host devbox (remote machines), but the docs describe no multi-user server or hosted service. If you are weighing a terminal factory against editor agents, read IDE agents vs terminal agents.
When something else fits better
GitHub Copilot cloud agent lives inside GitHub: you pick Copilot under an issue's Assignees, it works on a pull request and adds you as a reviewer when it finishes, and enterprise owners can review agent activity in the audit log (GitHub docs, audit log). Choose it if your team already runs on GitHub Enterprise and wants nothing to install. Devin is "an autonomous AI software engineer that can write, run and test code", taking work from Slack threads and Linear or Jira tickets (Devin docs). Choose it if you want a managed cloud service rather than a machine you run.
Questions people ask
How do I set up an AI software factory for my engineering team?
Start with one repo and one kind of work. With CodeAF: install it, write house rules with /standing, set budgets, hand off jobs with /task, then add a team with a manager and nightly schedules once the first jobs land cleanly.
What is an AI software factory?
It is a setup where agents size, split, run, check and merge work while people set goals and approve what lands. In CodeAF that means tasks on their own branches, managers over teams, and standing orders that keep running after you stop typing.
Will the agents push to main?
Not by themselves. CodeAF runs handed-off work in its own copy of your folder and keeps it on its own branch when you are on a branch such as main that tasks never merge into. senior-dev cuts a task branch when the folder has git history; in a plain folder it edits in place.
Is there a hosted CodeAF for the whole team?
No. CodeAF runs on machines you control, and the docs describe no hosted service. A team can share one dev box over ssh with codeaf chat --host devbox.
How do I stop the agents from overspending?
CodeAF prices every call before it runs, with a daily limit ($500 by default), a $5 per-task limit, team caps only you can raise, and ceilings on each specialist.
Sources
- CodeAF docs: FAQ, first team, managers, rules, budgets, fire and steer, specialists, schedules, watches, running unattended, build a software company, checked 2026-11-21.
- CodeAF repository, Apache 2.0, v0.8.3 (2026-11-19).
- GitHub Copilot: agents overview, audit logs, checked 2026-11-21.
- Devin: docs.devin.ai, checked 2026-11-21.
Last updated 2026-11-21. Checked against CodeAF v0.8.3 docs and the GitHub Copilot and Devin documentation on that date.