Why we built Tycho
- Published
- Reading
- 4 min read
- Written by
- NSi4
Chat assistants are good at snippets. We needed AI agents to change production systems: the engineering behind Trinity, our own tools, and nsi4.com itself. A chat window could not do that safely, so we built the place they work.
Tycho is our AI-native IDE. Agents plan, edit, build and test against real codebases in it, inside rails a person sets. This is the second post in our series on why each NSi4 product exists.
Chat assistants forget
A chat window loses its context when it closes. It runs every tool with the same permissions. It has no safe place to execute a command.
None of that matters for a snippet. All of it matters when the code runs a system that trades our own money.
What the agent must never lose
We started from what an agent has to keep, whatever happens to the machine it runs on.
- Its conversation. It lives in one remote session. The desktop app only projects it, so a crash or a restart resumes the same workspace instead of starting over.
- Its persona and its context. Every run composes a skill stack at its first message: the agent's persona, its tools and the project's context. A skill loaded later adds its tools from the next round.
- A safe place to run commands. The agent's terminal is an isolated PowerShell process, never the user's own shell, and every output stream is captured.
The rails
An agent in Tycho does not work from a chat prompt alone. It works from a task.
- Claim. The agent takes one task from the plan in AliveMD. The task says what to change, what the change must prove and how we will check it.
- Work. It runs in its own worktree, on its own branch. A guard keeps its tools on its own checkout, so two agents on one repository no longer break each other's builds.
- Prove. It runs the build and the tests, and for a website it captures the pages, in the terminal it drives.
- File. It writes the result and its checks back to the task.
- Verify. When the task completes or halts, Tycho rings the session that dispatched it. The work is checked against its task before it is accepted.
Every turn streams, runs its tools and is written to a trace, so any run can be read back afterwards. Tycho's own changes go one step further: our gate tests every one of them before it reaches the main branch.
It is building our website
The redesign of nsi4.com was executed by Tycho agents against tasks in AliveMD. On 8 October 2026 that plan counted 128 tasks executed by Tycho agents and 149 commits, each one checked against its task before it was accepted.
One of them, RD-96, edged the hub thumbnails with a hairline. The agent claimed the task, read the stylesheet, added seven lines and removed one, built a release of the site, captured the blog and research pages at desktop and phone widths, and filed its result with all three acceptance checks passed.
That is the shape of all of it: small, specified, proven and on record.
What runs it
Tycho is a desktop host with Blazor panels. Agents analyse code through Roslyn, with embeddings for semantic search, and work with guarded SQL, AliveMD documents and sandboxed files. AliveMD holds the plans, tasks and project context that Tycho's agents work from.
Tycho is in use at NSi4 and is not publicly available. If you would like it for your team, tell us what you would build with it at contact@nsi4.com.
Next in this series: AliveMD, the record every one of those tasks lives in.
More from the blog