← All work

17 / AI agent framework · solo · internal tool

harness-agent

A framework for letting an AI agent work for a long stretch without deciding the big things itself: at each important point it stops and waits for a person. Long-running agents burn money, drift, or do things nobody allowed, so the control sits in the framework and does not rely on the model behaving. NYX built it for its own use and does not sell it.

No public demo
A terminal running harness-agent on synthetic data: summary tables, two ranked rows, the cost of the run and an approval gate waiting for a person
harness-agent — synthetic data
Role
Solo — architecture and build
Status
Internal tool
Approval gates
10 kinds · 21 test files
Stack
Node 22 · TypeScript · SQLite · Claude Agent SDK · Vercel Sandbox

The rule

Limits are tripped by counters, never by what the agent says about itself.

Build notes

One way into the SDK

Nothing calls the agent SDK directly. A second path would mean a cost ledger that undercounts, approval gates that never appear, and per-role tool tables that turn into advice.

Agents cannot open a gate

Gates are opened by framework code around the loop, so there is no gate tool a manipulated model could reach. The 10 kinds are rows in one table. A late gate gets a louder reminder, and nothing ever decides for the person.

Foreign code in two phases

The sandbox has network while dependencies install and none before third-party code runs. That order lives inside the module, so it is a property of the system and not a habit of the caller. An agent with a sandbox has no outside credentials, and the reverse.

Hit a ceiling, stop

Assumptions, cost and time each have a ceiling, and each is a counter. On reaching one, the run stops first and opens a gate second. Nothing dead gets restarted automatically.

Want to work with NYX?

Build a product together, distribute one that exists, commission software, or join in — write a few lines and NYX will get back to you.

Send an invitation →