Book a demo

The self-healing layer for AI agents

Better agents.
Every day.

See every agent. Cut wasted spend.
Fix what breaks, automatically.

Book a demoSee how Newt works

Works where your agents already live

  • OpenAI Dots
  • Grok
  • Town
  • Muse
  • Viktor
  • Serval
  • Claude Projects
  • Notion Agents
  • Copilot Studio
  • Zapier
  • n8n

The problem

Why AI agents get worse over time

Your company already runs hundreds of AI agents. Quietly, they drift. They follow last quarter's instructions2025 travel policy, pull the wrong documentpricing_v2_DRAFT, hit tools they can't use403 forbidden, and get fixed by handre-run ×4.

Nobody owns making them better. Newt does.

How it works

One loop that keeps every agent improving.

  1. Find

    Finds what fails.

    Newt watches every run on every platform, and catches the re-runs, hand edits and dead ends people never report.

    pricing_answers
    Failing
    • 09:12Quote for Acme, 40 seatsok
    • 09:31Enterprise discount questionre-run ×3
    • 10:04Annual vs monthly pricingedited 61%
    • 10:22Renewal price for Globexok
    • 11:47Volume tier for 500 seatsre-run ×4

  2. Diagnose

    Figures out why.

    Each failure is traced to its root: instructions, retrieval, tools or permissions.

    pricing_answers
    Diagnosed

    Trace · Volume tier for 500 seats

    retrieved

    pricing_v2_DRAFT.pdf

    should have used

    pricing_2026.pdf

    root causeRetrieval

  3. Fix

    Tests a fix on real work.

    Newt drafts the change and replays real past tasks against it before anything ships.

    pricing_answers
    Testing

    - source: pricing/*

    + source: pricing_2026.pdf

    + answer with tier table

    Replaying 200 real tasks188 pass

  4. Prove

    Proves it worked.

    Every fix lands with evidence: fewer failures, fewer retries, fewer human corrections.

    pricing_answers
    Improved
    Success rate71%94%
    Retries3112
    Human corrections185

    Shipped · approved by Dana K.

Proof

Every fix ships with evidence.

You approve what changes. Newt shows exactly what got better, agent by agent.

Newt · this month

143 agents got better this month.

38 fixes shipped · 2 waiting for you

Failures
−48%
Retries
−62%
Human corrections
−71%

Needs your approval

  • policy_qaPoint to the 2026 travel policy96 / 100 replays pass
    Approve
  • crm.update_leadRequest the missing leads:write scope0 errors in 50 replays
    Approve

Success rate · before → after

  • pricing_answers71%94%
  • policy_qa64%88%
  • sales_recap58%90%
  • ticket_triage82%97%
  • onboarding_buddy77%91%

FAQ

Before you trust it with your agents.

Where do our agents need to live?

Wherever they already are. Newt connects to the platforms your teams build on, from ChatGPT and Claude to Notion, Copilot Studio, Zapier and n8n, and improves agents in place.

Does Newt change agents without asking?

No. Every fix is tested against real past tasks first, then sent to the agent owner with evidence. You decide what ships, or set rules for low-risk fixes to ship on their own.

Do we need an AI engineer to run it?

No. Newt was built for companies where agents are made by everyone, not by a dedicated AI team. It does the monitoring, diagnosis and testing for you.

How do you measure "better"?

Per agent, before and after every fix: how often tasks succeed, how often people re-run them, and how much of the output they rewrite by hand.