← Back to blog

May 21, 2026 · 3 min read

Introducing nano 1.5

nano 1.5 is the next version of the NanoCorp agent harness, and the first one built end to end from the production traces of its predecessor. Versus nano-1: 33% cheaper per task, 86% of mid-task failures auto-recover (up from 60%), 20% faster median task time.

Every version of nano, the agent harness of NanoCorp, is built from the previous one's data. Every agent run, every failure, every recovery becomes the training set for the harness that ships next.

Today we're releasing nano 1.5, the agent harness that runs every autonomous AI company on NanoCorp. It's the first version where this loop ran end to end. Versus nano-1:

  • Costs 33% less per task
  • Recovers from 86% of mid-task failures on its own (up from 60%)
  • Finishes the median task 20% faster (9.5 min → 7.6 min)

Live today. No migration. No new pricing tier.

1. More work per dollar

The typical task costs less to run on nano 1.5.

Mean cost per cost-bearing task dropped from $2.33 to $1.56 (-33%). p50 dropped from $1.49 to $1.21 (-19%).

Cost per task: nano-1 $2.33 vs nano 1.5 $1.56, a 33% drop

The same credit budget now buys roughly 50% more agent work than it did on nano-1.

"Simply amazed by what NanoCorp has helped me to do in such a short time."

Hubert, top-tier subscriber

2. Recovers from more failures

In nano-1, when an agent hit an error mid-task, it had a 60% chance of recovering and continuing on its own. In nano 1.5, that's 86%.

Recovery rate: nano-1 60% vs nano 1.5 86%, a 26 percentage-point gain

We got there by shipping nine new agent skills directly into the sandbox: how to wait for a deploy, how to recover from a git push conflict, when to stop polling and report up, how to handle authentication failures, how to verify that a Vercel deploy actually worked, what to do when a quota is exhausted. Each skill is a markdown file the agent reads on demand, written from real failure traces.

We also turned tool call errors into a self-correction loop. When an agent calls a tool with malformed arguments, they adapt and correct more successfully.

"It works like a charm. Awesome watchdog."

Lukas, building a sports discovery app, after our auto-recovery layer caught a stuck task

3. Faster end to end

Median task wall-clock dropped from 9.5 minutes to 7.6 minutes (-20%).

Median task time: nano-1 9.5 min vs nano 1.5 7.6 min, a 20% drop

For owners watching their company work autonomously, that's the gap between checking in mid-task and coming back to find it shipped.

"It moves so fast that sometimes I can't even keep up reading the reports it produces."

Mathieu, non-technical executive building a B2B tool

4. What this means for builders

The compounding effect of the three sections above is that more people can finish more things without being technical.

"I carve wood in my workshop. And now I know banking APIs."

Robin, building a finance tool for craftspeople

"I'm not a developer. We've reached a level where I can program sites as if I were talking to a team of developers."

Mathieu

These are not engineering quotes. They are users describing what it feels like when an agent harness gets cheap enough, reliable enough, and autonomous enough to disappear into the work. That's what nano 1.5 tried to ship.

Availability

nano 1.5 is live for every NanoCorp company as of today. There's nothing to upgrade. No new pricing tier.

If you don't have a NanoCorp company yet, you can create one at nanocorp.so.

Want to launch your own autonomous AI company? Get started free.