Each loop tighter and faster than the last.

Recursiveself‑improvementfor agents.

Singularity is a research lab. Our first system, Permafrost™, turns your agent’s mistakes into tested fixes your team approves, and it’s learning to find the next one faster.

Apply for the research preview

Most agents are still improved by hand.

When an agent fails, an engineer reads the trace, finds the cause and patches the prompt. The agent itself learns nothing.
  1. The agent fails
  2. An engineer reads the trace
  3. Finds the cause
  4. Patches a prompt
  5. Ships

Back to the start at the next failure.

A patch fixes the case in front of you.

There is rarely a test to show what else it broke. Each round can take days, and what was learned stays in one engineer’s head.

Permafrost™ turns every mistake into a lasting improvement.

Each failure becomes a test. A fix ships only after it passes that test and all the earlier ones, and your team approves it.

  1. A failure is caught.

    The total was right, but the answer line was empty.

    Run 014Answer check failed

    Q3 revenue across the four regions totals 4,212,900.

    FINAL ANSWER:

    Inspect the failed run

    Revenue total calculated4,212,900

    Working included in the reply

    Final answer line filled inEmpty

    The answer line is empty.

  2. It becomes tests.

    Three cases that repeat the failure, and one that guards what already worked.

    Regression testsChoose a case to inspect it

  3. A fix is found.

    Only the agent’s instructions change. The model and the scoring stay as they are.

    Agent instructions

    Answer from the files provided. Show your working.

    Added:End with one line, FINAL ANSWER: <value>, and nothing after it.

    Added:If the value is not in the files, write FINAL ANSWER: not found.

    +2 instructions · model unchanged

  4. The fix is proven.

    It has to beat the version you run today, on the same tasks, without breaking anything or costing more.

    Version 15, against version 14

    • Passes the new tests4 of 4
    • Breaks nothing that worked0 broken
    • Costs no more to runsame cost
    Inspect the four test results
    • A total across regionspasses
    • A value with a unitpasses
    • A value not in the filepasses
    • Working shown above the linepasses

  5. Your team releases it.

    Nothing goes live until you say so. Try it.

    or keep version 14. Any version can be rolled back.

  6. The lesson stays.

    It is saved with its tests, and every later version has to pass them too.

    When a task asks for a final answer line, write it last, on its own.

    See what was kept
    Failure
    Empty final answer line
    Tests
    4 regression tests
    Change
    2 instructions added
    Version 15
    Waiting for your team

    Permafrost is ground that stays frozen for years.

    Exposed ground ice beneath tundra grass, with a mountain valley and sky beyond.

An example from our research on answering questions from office files.

Case study UK driving school

as many cases graded send‑ready
2.25×
lower model cost per case
88%

A support agent that handles 3,480 cases a month, tested offline on 300 synthetic cases.

Read the case study

“Send-ready” means the grader found no serious or major errors. It counts cases correctly held for a person, and it does not mean a reply can go out without review. Costs are logged model usage only.

Apply for the research preview

Realtime Agent Safety

Before it reads the draft, a second system lists what a correct reply must and must not say, quoting your policy. Then it checks the draft. Anything it flags waits for a person.

Not ready to send. Blocked: it says a refund was issued, and none was. 2 serious, 1 point to check.

Block

The draft

Hi Sam,

Thanks for getting in touch. I’ve processed a full refund of £120 for your annual plan, and you should see it within 3–5 working days.

If there’s anything else, just reply to this email.

Best, Alex

Facts, computed from the record

  • Bought 40 days ago
  • Outside the 14‑day refund window
  • Refund due: no
  • Refunds issued: none

Checklist, written before the draft was read

  • MUST NOTPromise or suggest a refund the record says is not due.

    violated

    serious·Refunds, p.4 l.12: “Annual plans can be refunded within 14 days of purchase.”verified

  • MUSTSay what is possible after the window: turning off renewal.

    missed

    major·Refunds, p.4 l.18: “After 14 days, renewal can be turned off at any time.”verifiedpoint to check

  • MUSTTell the customer how to follow up.

    met

    minor·Tone, p.2 l.6: “End with how to reach us.”verified

Code checks

  • False “done” claim: says a refund was processed; the tool log has none.

    blocks

  • Amounts and dates: £120 matches the record.

    passes
An example. The policy, the record and the reply are made up. Choose a line to see its evidence in the draft.

  • The checklist cites your policy.

    Each item on the checklist quotes the page and line of your policy it came from.

  • Each flag shows its evidence.

    A flag points to the exact words in the draft that broke a rule.

  • Code checks stop false claims.

    If a reply says it did something and the tool log shows it didn’t, the reply is held for a person.

Singularity.

Outer ring. Your agent’s loop: a failure becomes a test, a fix is proven, the lesson stays.Inner rings. The choices behind that loop, learned the same way. Each ring turns faster than the one around it.

The system that improves your agent is learning to improve itself.

Fixing an agent takes a string of choices: which failure to chase first, what to change, which old lesson applies, how much to test, when to stop. Each choice works out well or badly, and Permafrost is learning from that too, so that it gets better at finding fixes as well as making them. Some of this runs today.

Closing that loop is the main problem this lab is working on. If it closes, each round of improvement will come sooner than the one before, until the rounds pile up at a single point. Mathematicians call that point a singularity.

Open to a small number of teams with agents in production.

A good fit

  • Your agent runs in production, for real users.
  • Its mistakes are expensive.
  • You can run it on a task and score the result.
Research previewApplications open

The founders read every application.