hawkes.aiKey ideas
- The one idea — A coding agent is a concrete, powerful harness: it wraps a general model with file tools, a shell, codebase search, a think-act-observe loop, context management, and objective verification (run the tests) — everything from these acts, composed.
- How it is shown — A coding agent reading a file, editing it, running tests, reading the errors, and fixing — the agent loop over a real dev environment, with tests as objective checks.
- The trap to avoid — Thinking a coding agent is a fundamentally different/special model — it's the same general model wrapped in a harness with the right tools, loop, context management, and checks.
- What it sets up — Inside all this machinery, what does the model actually see?
The AI that edits your files and runs your commands isn't a special coding model. It's a general model in a great harness — and you already know every part.
The one idea
A coding agent is a concrete, powerful harness: it wraps a general model with file tools, a shell, codebase search, a think-act-observe loop, context management, and objective verification (run the tests) — everything from these acts, composed.
Let's make the harness concrete with the most powerful example: a coding agent. It reads your files, edits them, runs commands, and works a task for many steps. It looks like a special kind of AI. It isn't — it's a general model in a very good harness, and its parts are all things you've met. Start with its tools. It can read a file, write and edit files, run shell commands, search the codebase, run the tests. The model emits the calls; the harness executes them — the tool calling you know, now pointed at real code. The tools are ordinary; wiring them to code is what makes it powerful. Then the loop. Think — I need this file. Act — read it.
How it works — the demo
A coding agent reading a file, editing it, running tests, reading the errors, and fixing — the agent loop over a real dev environment, with tests as objective checks.
Observe — the contents. Think again — edit it. Run the tests. Read the errors. Fix. It repeats that cycle autonomously, until the task is done. It's the agent loop from the last act, doing real development work, one concrete step at a time. Now context. It can't fit your whole codebase in the window, so it doesn't try. It opens only the relevant files, searches for what it needs, and summarizes as it goes — curating context to fight the rot. It's retrieval applied to code: give the model the right slice, not the whole ocean.
The trap to avoid
Thinking a coding agent is a fundamentally different/special model — it's the same general model wrapped in a harness with the right tools, loop, context management, and checks.
Why it matters — and what’s next
Inside all this machinery, what does the model actually see?
And it checks itself. It runs the tests and the compiler — objective checks, the world answering, not the model's opinion. If they fail, it reads the error and tries again. Verification is wired into the loop, not left to hope. That one habit — run the tests — is much of why a good coding agent is trustworthy. The trap: thinking a coding agent is a fundamentally different, special kind of model. It's the same general model you've met all series, wrapped in a harness with the right tools, loop, context management, and checks — everything from these acts, composed. Change the harness, not the model, and you get a different agent. So a coding agent is a general model in a great harness — tools, loop, context, verification, all wired together. But here's a question that dissolves a lot of confusion: inside all this machinery, past the interface and the files and the tools, what does the model actually see? Next: what the model actually sees.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Let's make the harness concrete with the most powerful example: a coding agent. It reads your files, edits them, runs commands, and works a task for many steps. It looks like a special kind of AI. It isn't — it's a general model in a very good harness, and its parts are all things you've met.
Start with its tools. It can read a file, write and edit files, run shell commands, search the codebase, run the tests. The model emits the calls; the harness executes them — the tool calling you know, now pointed at real code. The tools are ordinary; wiring them to code is what makes it powerful.
Then the loop. Think — I need this file. Act — read it. Observe — the contents. Think again — edit it. Run the tests. Read the errors. Fix. It repeats that cycle autonomously, until the task is done. It's the agent loop from the last act, doing real development work, one concrete step at a time.
Now context. It can't fit your whole codebase in the window, so it doesn't try. It opens only the relevant files, searches for what it needs, and summarizes as it goes — curating context to fight the rot. It's retrieval applied to code: give the model the right slice, not the whole ocean.
And it checks itself. It runs the tests and the compiler — objective checks, the world answering, not the model's opinion. If they fail, it reads the error and tries again. Verification is wired into the loop, not left to hope. That one habit — run the tests — is much of why a good coding agent is trustworthy.
The trap: thinking a coding agent is a fundamentally different, special kind of model. It's the same general model you've met all series, wrapped in a harness with the right tools, loop, context management, and checks — everything from these acts, composed. Change the harness, not the model, and you get a different agent.
So a coding agent is a general model in a great harness — tools, loop, context, verification, all wired together. But here's a question that dissolves a lot of confusion: inside all this machinery, past the interface and the files and the tools, what does the model actually see? Next: what the model actually sees.