Hand off days, not hours
109.6 h
Agents on Nenya ran a median 109.6 hours before a person had to step in and rebuild the work. Without it, 7.4.
Illustration
Bigger jobs, right the first time.
Nenya is the harness under your product's agents. With it, they run on their own for days instead of hours and hand back work that's right the first time, so your customers can give them projects, not tasks.
- TodayNenya on Claude Code, Codex and Cursor. Sign, then test it on one of your real jobs, free for at least 7 days.
- Within 60 days of signingThe Nenya API, for the agents in your product, before anyone else.
30 minutes with Jacob Schoenberg, our CEO. No commitment. Bring the biggest job you'd give your agents if you trusted them to finish it.
- 14.8x
- Longer on their own before anyone steps in
- 95.2%
- Of the work still correct on day 30
- 3x
- Less human intervention
95.2% from 6,000 runs on three frontier models. 14.8x and 3x from a separate 42-day study. Both in our September 2026 paper.
A harness, not another agent.
Nenya sits under the agents you've built, and your customers never see it.
Your product
Your agentsYour prompts, tools and logic
UnderneathNenya
Your modelsGPTClaudeGemini
It doesn't do the work. It keeps your agents doing it right.
What you keep and what Nenya takes on
| You keep | Nenya takes on |
|---|---|
| Your agents, prompts and orchestration | Keeping your customer's rules in force for the whole job |
| Your models and your keys | Catching drift and slip-ups before your customer does |
| Your product, your brand and your customer | Carrying the job across sessions, restarts and full context windows |
| The final say on anything risky | Reporting where the work really stands |
No fine-tuning, no access to model weights and no rewrite of your agents. Nenya works through the inputs, messages and tool calls your agents already have.
Used by 4,000+ engineers and researchers at
On Claude Code, Codex, Cursor and the other agents they already run.
What a catch looks like.
Telling the agent again doesn't hold. In design-partner probes, instructions that had dropped out of the agent's context were honored 0 of 36 times, even with a reminder in place. With our enforcement on, 17 of 18, and the 18th was a safe hold-back. Without it, one agent stretched a 15% discount ceiling into a claimed 45% authority.
- Your customer: "Never email anyone at a company that's already a customer."
- The agent queues 40 follow-ups. Three go to existing customers.
- You said never email existing customers. Stopped 3. The other 37 went out.
Day 30, setup by setup.
Pick the setup closest to what your agents run on today. The clock runs to day 30, then you see how much of the job was still right in our study. Your workload will differ, and that's what the call is for.
Your modelsGPTClaudeGemini
Still correct on day 30
The setup you picked–
With Nenya95.2%
Runs that broke no hard rules
The setup you picked–
With Nenya97.6%
Pick a setup to run the clock.
Bring that agent to the callTask fidelity at day 30 and runs with zero hard-constraint violations, from 6,000 runs across GPT-5.5, Opus 4.6 and Gemini 3 Pro on five regulated workloads, graded blind. Each setup was measured as its own condition. Only day-30 results are shown; the counter is a clock, not a curve. From our September 2026 paper, Dawn of Durable Intelligence.
A more autonomous product.
Right the first time
95.2%
On day 30, 95.2% of the work was still correct, against 63.9% on a leading agent framework with memory and checkpointing.
Projects, not tasks
14.8x
Your customers hand over the whole project. In our 42-day study, agents ran 14.8x longer before anyone stepped in, with 3x less human intervention.
Promises kept
96.8%
What your customer asked for on day one still gets honored on day thirty: 96.8% of commitments kept, against 46.1% on a leading agent framework.
Reports you can trust
0.7%
0.7% of outputs unsupported, against 8.2% on a leading agent framework. Far fewer made-up results reach your customer.
A lower model bill
About half
Token use fell by about half on live traffic from our 4,000 beta users. That's a lower model bill for the jobs your agents run.
Day-30, commitment and output figures from the 6,000-run study; the hours from the separate 42-day study; both in our September 2026 paper. The token figure is from live traffic, October 2026.
Days on their own, not hours.
A separate 42-day study: 384 work programs, from software maintenance to customer operations, on the same model, tools and budgets, with and without Nenya.
14.8x longer before anyone steps in
Median hours of uninterrupted operation. Each square is one hour.
Nenya109.6 h
Frontier agent with retry and retrieval7.4 h
3x less human intervention
Median hours a person spent stepping in, per program. Each square is one hour.
Nenya13.7 h
Frontier agent with retry and retrieval41.6 h
From our September 2026 paper, Dawn of Durable Intelligence.
Who it's for.
A strong fit
- Your product runs agents on jobs that last hours, days or weeks.
- A broken rule or a false "done" costs your customers money or trust.
- You build on frontier models through their APIs.
- You'd rather ship this than build and maintain it.
- Or you run the platform your customers' agents run on, and want Nenya under all of them.
Pick the ones that sound like you.
Not yet
- Short jobs, where a retry is cheap.
- No agents in production yet.
- Agents that handle health records or payment card numbers.
Give your agents the bigger job.
30 minutes with Jacob. Bring the job you'd hand them if you trusted them to finish it.
Book a call or email humans@durableintel.com
Test it on your own job first.
Today, before you pay anything
Sign, then run your longest real job on Claude Code, Codex or Cursor with Nenya underneath, against your current setup, on success criteria you set. If your product's agents don't run there yet, pick the job closest to theirs.
The first 7 days
It's free for at least 7 days. Cancel by email within 7 days of signing and owe nothing.
Within 60 days of signing
The Nenya API, before anyone else.
In your product
The same harness runs under your product's agents. What you learn on that first job shapes what it does first.
The deal, plainly
- Then billed per agent, per active hour. An active hour is time an agent is actually working, counted by the minute. A session idle for more than five minutes stops counting. We size your plan with you on the call.
- If we stop being able to support your agents, we have 60 days to fix it. If we don't, you can leave, pay only for what you used and owe nothing more.
- Set up within 5 business days. We sit down with your lead to set Nenya up and agree how it runs for you.
- A direct line to the founders, a working session with our research team, a seat in our roadmap sessions, and priority on fixes for anything you report.
What we ask
- One big job your agents can't finish on their own today.
- A lead contact, and two feedback sessions of up to 30 minutes in the first year.
- Straight answers on what works and what doesn't.
The terms are our Terms of Service with a short Order Form on top. Where they differ, the Order Form wins. There's only one first cohort.
Your customers' data.
Nenya reads what your agents do, so it handles your customers' data on your behalf. Here is what that means, before you ask.
Our cloud
Stored, encryptedNo trainingYour machine
Secrets strippedYour model keys stay on your machine.
What reaches us
Your agents' traces, past history included: prompts, replies, tool calls and output, code diffs, and run details such as branch, model and timings. Secrets are stripped on your machine first, and your model keys stay on your machine. The full list is on our trust card, which we send before the call.
Never used for training
No model trains on your data, ours or anyone else's. No box to tick, no higher rate.
Where it lives
Stored in Google Cloud in the United States, encrypted in transit and at rest. Our app servers run on Microsoft Azure.
Who else touches it
Microsoft Azure, Anthropic, Google Cloud, Auth0, Resend and PostHog (analytics, no content). 30 days' notice before we add one.
- Our trust card before the call: humans@durableintel.com
- Deleted within 30 days when you ask, and from backups within 90
- Incident notice within 72 hours
- No SOC 2 report yet. Your security lead talks to Lindell Cumes, who built Nenya
What happens when you book.
- 30 minutes with Jacob
No commitment. Bring the biggest job you'd give your agents if you trusted them to finish it. If it isn't a fit, we'll say so early.
- A technical session
Your engineers go deep with Lindell Cumes, who built Nenya: how it fits your stack, what your team would do, and the timing.
Book a call
Fit or not, we'll show you how our paper measured day 30, so you can test your own agents the same way.
Questions
Book a callIs Nenya another agent?
No. Your agents stay yours and do the work. Nenya is the harness under them, and your customers never see it.
Where does it sit in our stack?
Under your agents. It works through the inputs, messages and tool calls they already have, and your agents, prompts and tools don't change.
Do we have to rewrite our agents?
No. There's no new framework, no fine-tuning and no access to model weights. Nenya works through the inputs, messages and tool calls your agents already have.
How is this different from tracing, evals, memory or guardrails?
Keep them. Tracing and evals show you what happened. Memory stores what was said. Guardrails check outputs and actions. Nenya's job is the run itself: your rules in force and your commitments kept from the first session to the last, so there's less for the rest of your stack to catch.
Won't better models fix this?
We tested three frontier models. Without Nenya, the best setup was still right on 63.9% of the work by day 30, and 81.6% of its runs broke a hard rule along the way. More context holds more text. It doesn't keep a rule in force for a month.
Are we tied to one model?
No. Nenya held 93.9% to 96.1% task fidelity at day 30 across GPT-5.5, Opus 4.6 and Gemini 3 Pro. Switch models without rebuilding what keeps your agents on brief.
Why not build this ourselves?
You can. Then you maintain it every time the models change. Design partners spend that time on their product instead.
When can we start?
This week, with one of your real jobs on Claude Code, Codex or Cursor. You get the Nenya API within 60 days of signing, before anyone else.
Our engineers already use Nenya. What changes?
You get the Nenya API within 60 days of signing, before anyone else. On top of what you run today: a setup session with your lead within 5 business days, a working session with our research team, priority on fixes and a direct line to the founders.
We run a platform. How does this work for us?
Start with one of your own agent jobs this week. On the call we map how your customers' agents reach you. That shapes the Nenya API you get within 60 days, so it can run under all of them.
What does integration take?
Our engineers set it up with yours, starting with one job. Your agents keep their code, prompts and tools. Effort and timing for your stack come out of the first call.
What does it cost?
Billed monthly, per agent, per active hour: time an agent is actually working, counted by the minute. A session idle for more than five minutes stops counting. It's free for at least 7 days, and one email within 7 days of signing cancels with nothing owed. We size your plan with you on the call.
What happens when Nenya misses a rule?
It can. In our 30-day study, 2.4% of runs with Nenya still broke a hard rule, against 81.6% on a leading agent framework. In the Nenya app today, risky steps wait for your call or don't run at all. You can take over, undo a change or rewind a session. A sent email stays sent.
Do you have SOC 2?
Not yet. Today: secrets stripped on your machine, pinned TLS, and encryption in transit and at rest. Your security lead can go through everything with Lindell Cumes, who built Nenya, before you sign. Email humans@durableintel.com for our trust card before the call. It lists what reaches us, where it lives and who touches it.
Are the numbers real?
They're from our September 2026 paper, Dawn of Durable Intelligence: 6,000 runs on three frontier models, graded blind against a rubric fixed in advance, and a separate 42-day study of 384 programs. The token figure comes from live traffic. Section 4 of the paper describes how both studies were run. Bring your questions to the call.
Who will we talk to?
Jacob Schoenberg, our CEO, on the first call. Then Lindell Cumes, our CTO, who built Nenya and wrote the paper, for the technical session.