Early failure detection for AI agents

Stop paying
for agent runs
that fail.

First Light detects agent runs that would eventually fail and course-corrects them early, saving you the spend and the hours spent waiting.

Send us a sample of your agent's traces How it works
  • Pilots underway
  • 20,441 agent runs audited
  • Founded by quantitative finance and astrophysics researchers
Example run
Without First Light
✕ failed
Ran all 12 steps, then failed.
With First Light
Flagged
Flagged at step 4. Repaired, then resolved.
Signal

The same test fails after three different patches, and edits are drifting into files the task never touches.

Course-correct: repair

Re-scope to the failing module and retry with the full error trace.

01 / The problem

Half of agent spend goes to runs that fail.

Across 20,441 public coding-agent runs, 52.9% of the money went to runs that never succeeded. Teams pay twice: once in compute, and again in the hours spent waiting for a result that was never coming.

52.9%

of agent spend went to runs that failed

1.5×

what a failed run costs compared with a successful one

873K

model calls analyzed in our audit

What we found

Generic warning signs don't work.

8 in 10runs that hit failing tests or tool errors still went on to succeed

Kill switches and step caps throw away runs that would have worked. The signals that predict failure are specific to each workflow.

So they have to be learned from your own runs. That is what First Light does.

Why nowAgents are moving to outcome-based pricing. When you charge per outcome, every failed run comes out of your margin.
02 / How it works

Early failure detection, trained on your runs.

First Light learns what failure looks like in your workflow, watches every run, and course-corrects the ones heading for failure while there is still time to save them.

  1. 01

    Measure

    We compute your real cost and time per successful outcome from your traces. Reported agent cost can swing by up to 455% depending on how it is counted, so most teams have never seen this number.

  2. 02

    Learn

    We learn the early signals of failure from your own successes and failures. Not generic rules: your workflow and your definition of done.

  3. 03

    Detect

    Every live run is scored as it goes. Our detectors are distilled into small, fast models, so watching a run costs a fraction of the run.

  4. 04

    Course-correct

    When a run is heading for failure, First Light stops it, repairs it or reroutes it. Every outcome sharpens the next prediction.

Drops into your stack

Point your agent's API base URL at First Light, or send us OpenTelemetry traces. No rewrite of your agent.

Where it fits

Observability tells you what happened. Evals tell you what went wrong. First Light changes what happens next.

Built for agent teams with long, paid, multi-step runs and a clear definition of done: coding, browser, voice and generative agents.

03 / Team

We find signal in noise.

We find weak signals in massive, noisy data, from stochastic risk modeling and tail-event prediction in financial markets to complex astrophysical datasets. Now we're applying that to AI agents.

Both founders trained in quantitative finance at Stony Brook University under former Renaissance Technologies researchers.

Aman Arya

Co-founder LinkedIn

Evaluates AI-generated code for Microsoft Copilot. Built machine learning models on large astrophysical datasets at the Indian Institute of Astrophysics, with peer-reviewed research published in the Journal of Astrophysics and Astronomy. MS in Data Science from Stony Brook, with graduate coursework in computational finance, stochastic models and machine learning for quantitative finance. Author of the SWE-bench agent spend audit.

Kunal Gaurav, CQF

Co-founder LinkedIn

Quantitative researcher at a hedge fund and an investment firm, building stochastic volatility, statistical arbitrage and portfolio risk models. Then senior data scientist, and now in AI operations research at Uber AI Solutions. Published in the International Journal of Financial Engineering on stress testing and tail risk. Previously founded a data and analytics venture. Studied at Stony Brook University.

04 / Design partners

Show us your traces. We'll show you what failure costs.

We work with a small number of agent teams at a time. We ask for two things.

01

Your definition of success and failure, in writing.

For each workflow, what counts as a successful run (an approved demo, a merged PR, a resolved ticket) and what counts as a failure (rejected, re-rendered, reverted, escalated, abandoned). Plus what a wrongly stopped run costs you. We agree on this before we see any data, and every result we report is scored against it.

02

Traces of your agents.

Past runs with every step and its outcome.

  • For each step: its type (model call, tool call, render), timestamps, model, tokens or cost, and status or error.
  • For each run: the final outcome, plus the reviewer's decision and reason if you have one.

OpenTelemetry, LangSmith or Langfuse exports, or plain JSONL. Redact what you need to.

What you get back

In 48 hours

Your real cost and time per successful outcome, and where the spend on failed runs goes.

Within 3 weeks

A shadow backtest on your own runs: which ones First Light would have course-corrected, what it would have saved, scored against your definitions.

Then

A paid pilot priced on measured savings, once we hit the success bar we agreed up front.

Your data: read-only, under NDA, never shared with another customer, deleted on request with written confirmation. We can run inside your cloud.

Send us a sample of your agent's traces.

We'll send back your cost per successful outcome in 48 hours.

Start with a sample
Research

The data behind every number on this page.

Our audit, methods, experiments and what they taught us.