firstresponse

Tech: Europe Agentic AI Hack / London, 19 Sep 2026

Evidence before action.

First Response is your voice incident copilot. Stay in the conversation while it finds the cause, generates repairs, and tests the code. You make the final call.

Explore the source

From failure to verified recovery

02:00 / 1080p

Edited highlights from real runs. Scenes 1–7 use synthesized narration; the final scene includes an actual Gemini Live response. Captions are included. Recording details

Download video
Generated candidates
3independent repairs
Checks per candidate
25reproduction + regressions
Recovery probes
12healthy checkouts after approval
Time to verified candidates
10.0seconds in the recorded timeout run

01 / The proof

A plausible fix
can still be wrong.

A configuration change cut the payment timeout from 3,000 milliseconds to three. Checkout failed. Gemini proposed three repairs from the broken source, without seeing a library of fixes.

All three fixed the immediate failure. An unchanged regression test caught what the five-second repair missed.

Read the diffs and test logs
Actual Modal results · payment timeout incident
Candidate Reproduction All checks
3,000 ms Original budget Pass 25 / 25Approved & rechecked
1,000 ms Shorter budget Pass 25 / 25Verified alternative
5,000 ms Longer budget Pass 24 / 25Rejected
Failed regressiontest_provider_timeout_remains_enforced

The five-second repair no longer raised the expected PaymentTimeout. Passing the reproduction alone wasn’t enough.

02 / How it works

A conversation up front.
An investigation in parallel.

Full architecture

Gemini Live carries the conversation. A FastAPI hub gives it incident facts while a separate Pydantic AI agent investigates and generates repairs.

  1. 01

    Observe the failure

    Real checkout probes, a stack trace, the deploy diff, and numbered source ground the diagnosis.

    Python shop · Pydantic AI
  2. 02

    Generate three edits

    Gemini proposes distinct, bounded source edits. Each is tied to the exact source snapshot.

    Gemini · validated output
  3. 03

    Test independently

    Each candidate runs in its own network-disabled sandbox against a reproduction and 24 unchanged regressions.

    Modal · parallel sandboxes
  4. 04

    Approve & recover

    The operator confirms a verified repair. Fresh Modal tests and checkout probes check recovery before a patch is ready to download.

    Human approval · fresh proof
Gemini Live

Ask about the cause or the failed candidate. Interrupt and keep talking. In the tested stream, a spoken apply request opens patch review; a typed command or button confirms the repair.

03 / Built to be inspected

Two real faults.
One complete repair loop.

The second incident removes a guard for an optional coupon. Gemini generated three distinct repairs; all three passed. Both incidents reached verified recovery through the same workflow.

This prototype covers two controlled faults in a toy shop. Generated edits are limited to one existing assignment line, tests remain independent, and generated code executes only in Modal. The hosted page is a demo showcase; run the application from the repository.

Voice checks used synthesized microphone input with the real Gemini relay. Human headset rehearsal remains separate. Timing is a single observed run, not a performance guarantee.

Run it locally Final integration evidence
Read the demo transcript