Introducing Naiad Agent: coding work that can keep going
Most coding assistants are strongest in the moment: explain this file, write this function, fix this error.
That is useful, but larger software tasks often do not fit inside one prompt. A task can take hours. It can involve several files, a failing test, a half-right first patch, and the need to come back later without losing the thread.
Naiad Agent is built around that gap.
Why another coding agent?
I wanted an agent that treated software work less like a chat response and more like a work session.
The goal is simple: give it a concrete task, let it work through the repo, and make the path visible enough that you can trust what happened.
Naiad Agent is built around:
- checkpoints, so work can be resumed or reviewed without losing context
- GitHub-native tasks, so the output can become a pull request
- visible traces, so you can see commands, edits, tests, and decisions
- longer-running sessions, so the agent can keep going past the first reply
Agent still leaves judgment with the engineer. It handles more of the legwork around a well-scoped task than a normal coding chat can.
What it feels like
A good Agent task is specific:
CSV invoice exports are double-counting discounted line items. Fix the totals, add coverage, and open a PR.
From there, Agent can inspect the codebase, find the relevant files, make a patch, run tests, and summarize what changed. The patch matters, and so does the trail behind it.
You can see what it did. You can stop it. You can pick the thread back up.
That is the shape I want from coding agents: less magic, more working memory, and a clear record of what happened.
How it fits with Naiad Lens
Naiad Lens helps you understand a codebase visually.
Naiad Agent helps you move a software task forward.
They come from the same belief: developers need AI that can work with the shape of a codebase instead of isolated snippets.
Lens is for maps. Agent is for motion.
Where it is now
Naiad Agent is available at v2.naiadai.com.
It is still early, and much of the work is driven by internal use. The product is strongest when the task has a clear goal, a repo to inspect, and a concrete definition of done.
That is where coding agents start to earn their place: carrying bounded engineering work through to a reviewable result.
-Eric