When a coding agent builds your AI agent, what should it build on?
Let the coding agent write what only your agent does: the tools that call your systems, the agent’s instructions and your business rules. Have it build on something that already does what every agent needs in production: approvals before risky actions, a record of every step, versions you can roll back, and tests on real cases before a change ships. Pick a stack it can read as it works, with instructions for coding agents that match the version you installed and errors that say what’s wrong.
A coding agent is an AI assistant that reads, writes and runs code in your project, much as a developer would.
What should the coding agent write itself?
Your logic. The tools that read from and write to your systems, like an order lookup or a refund. The agent’s instructions. The rules only your business has, such as a refund limit or who gets an angry customer’s ticket. These change when your business changes, and they’re the parts you’d want to read in review anyway.
For example: for Acme’s order-desk agent, the coding agent writes the order lookup that calls Acme’s store, a refund tool that stops at Acme’s $100 limit, and the instructions for answering an unhappy customer. The approval before a bigger refund, and the record of who gave it, come from what it builds on.
What should the coding agent build on instead?
The parts every agent needs, which have to keep working the same way for months: approvals, the record of what ran, versions and rollback, tests before a release, and keys kept out of code. A coding agent can write each of these in an afternoon. But every project then gets its own slightly different version, and each one is code your team now reviews, patches and explains to an auditor. We wrote more on that trade-off in Build vs buy for AI agents.
Should every step be the agent’s decision?
No. Where the order is the rule, write it as a fixed flow: check the order, then refund, then send the email. Use an agent step where judgment is needed, like reading what a customer actually wants. A fixed flow is easier to test, and easier to explain when someone asks why something happened. More in Workflow automation or an AI agent.
Is the model’s explanation an audit trail?
No. An audit trail is the record of what actually ran: each model call, each tool call with its arguments, what came back, and who approved what. The model’s account of its own reasoning is a claim about what it did, not a record of it. Build on something that writes the record for you, before each step moves on.
How does a coding agent learn a framework it hasn’t seen?
From what the framework gives it to read, not from memory. A coding agent that half-remembers an older version writes code against an API you don’t have. Look for three things: instructions written for coding agents that match the version you installed, docs it can fetch as plain text, and error messages that say what’s wrong, so it can fix the code and run it again.
Kindgi does all three, and more. Each project gets Kindgi’s skills, instructions written for coding agents that match the version it installed and refresh after an upgrade. Kindgi’s code and its commands explain themselves, every docs page has a Markdown version, and errors name what’s wrong. The coding agent also runs what it wrote, reads the record of each step to see what happened, and fixes it. If the problem turns out to be in Kindgi itself, it writes up the bug for us, and sends it only after you say yes. Later, it checks back and tells you when it’s fixed, and in which version.
What should stay with you, not the coding agent?
The names, the keys, what the code can reach, and the yes before anything risky. Your model key should never pass through the coding agent’s chat. The code it writes should run where it can’t read your keys or your other projects.
Where Kindgi fits
Kindgi is an agent framework for production: your coding agent writes the tools, agents and flows in your own code, and Kindgi runs them and keeps every step on record. To start, you paste one prompt into your coding agent: “Set up Kindgi for me: read https://docs.kindgi.com/start/agent.md and follow it step by step. Tell me whenever you need me to do something, and wait for me.” It checks your machine, makes the project, starts Kindgi on your machine and writes a first agent. When it needs your model key, the key goes from your clipboard straight into a command, never into the chat.
Your coding agent writes in TypeScript or Python, and the code runs sandboxed, away from your keys and your other projects. Approvals, the journal of every step, versions and rollback are already there.
The guides: set up Kindgi with your coding agent, your coding agent, send us feedback, ask before a tool runs and read a run’s journal.
How this was written: a coding agent drafted it from our docs and our code. We checked every claim, edited it and said yes. We publish articles like this partly because we want you to find us, including through AI assistants. More in How we write our posts.