14 September 2026
Heard In AI

Before the prompt: Peregrine says agents write about 90% of its integration notebooks

On the Training Data podcast, Peregrine's Ben Rudolph described integration agents that run for hours, inspect a customer's databases and split work among sub-agents, writing roughly 90% of the Python notebooks the company uses to connect public-safety records, under the deployment team's oversight. The conversation put the share of effort that happens before a user types a question at 95% — the preparation that let a Florida county ask why it had suddenly run more than a hundred water rescues.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

A county in Florida had recently carried out more than a hundred water rescues — crews taking boats out to people stranded in flooded cars and homes — and nobody could say why the number was so high. On the Training Data podcast, Peregrine's Ben Rudolph described what happened when the agency put that question to the company's software: after a few rounds of back-and-forth and some deep research by the agent, it surfaced a pattern in the weather conditions, which the county had seen before but never across three consecutive days. Those conditions, as the analysis went, carve sand channels, and sand channels make ideal conditions for rip currents, which cause trouble for people in the water.

Rudolph presented this as something the agency could then examine itself — overlaying its terrain maps and its own county systems — rather than as a settled explanation handed down by a model. The floods and the rip currents are not the same hazard, and he did not claim the agent had proved one caused the rescues. What he did claim is that the question could be asked at all: "there's zero chance we would have figured that out," he said of the idea that Peregrine's own staff might have guessed from headquarters what emergency responders in a hurricane needed.

Behind that answer sat incident reports, 911 calls and weather data, all integrated and available to the AI. And that, on the podcast, is where nearly all of the work was said to live. "Certainly 95% of the work is what happens before the user types in the question," as Nick Noone put the share — the preparation required for the system to answer accurately and cite its sources correctly, which the founders called a genuinely hard problem and a large part of where their engineering effort goes.

What has to happen first

Peregrine sells software to police departments, fire services and emergency agencies, and its starting position is that the data belongs to the customer. The platform's job is to connect systems an agency already owns so that records scattered across separate databases can be queried together.

That connecting work is more than copying tables. In a March essay on operational AI, Rudolph describes building a semantic ontology: a structured model that maps raw records onto meaningful entities — people, places, vehicles, incidents — and the relationships between them, with databases, documents and rich media ingested into that shared representation. His example from outside public safety is an airline, where weather, maintenance and scheduling information have to be joined before an agent can tell a tolerable maintenance issue from one that grounds an aircraft.

Local meaning has to be encoded too. The same essay describes engineers learning that a department uses "graves" for overnight shifts, or that a weekly briefing should land on Sunday evening rather than Monday morning, and building those conventions into the deployment.

Rudolph put roughly half of Peregrine's engineers on the data platform that lets the deployment team carry out this integration work, and said the company has integrated tens of thousands of datasets across its customers.

Agents doing the unglamorous part

The integrations themselves are written in a version of Python notebooks — documents that mix code with the results of running it, useful for inspecting and transforming data step by step. Peregrine built an agent to write them.

"About 90% of that is written by agents with the oversight of our deployment team," Rudolph said. These runs are long: the agent analyzes the customer's databases, works out the ontology, and starts piecing together what needs to be integrated, splitting the work into sub-agents that report back to an orchestrating agent. "These agents run for hours," he said. The host compared it to engineers pointing coding agents at codebase migrations — long-running, unglamorous work — and Rudolph agreed.

He also gave a reason why this particular job suits agents. "I love this problem because it's verifiable, and that makes the problem a lot easier," he said, adding that this is part of why coding agents work as well as they do. Peregrine has built an evaluation set against which it can deterministically check an integration agent's output for completeness and correctness — whether everything that should have been brought across was brought across, and whether it is right — rather than relying on a human judgment call about whether the result looks plausible.

Rudolph framed the underlying strategy as unchanged by AI. A chart can be useless if it is drawing on the wrong or inaccurate information, he said, and the same is true of agents, and will be true of whatever technology comes next; most of it is downstream of high-quality data.

What people ask once the records are joined

When the software first arrives, Rudolph said, what an agency gets is "a nice search": look up an address and see everything that happened there, formatted readably, instead of scrolling rows. What interests him is what comes after users understand the system, when the floor rises across a whole department and analyses appear that he describes as previously impossible.

His other example was a detective investigating a threat against a synagogue, who wanted to know whether two synagogues in the area had received other antisemitic threats. Rudolph's point was about the mechanics of searching: it is very hard to guess which keywords would find those reports. Working from meaning rather than exact wording, the detective was able to pull out a pattern of threats against the synagogues.

Peregrine's investigation assistant, described on the company's product page as in early access with selected agencies, works on the same principle from the other direction: it cross-references material from a case against the agency records a user is authorized to see — finding a phone number that appears in otherwise unrelated cases, or connecting an address to earlier weapons calls — and returns citations back to the underlying evidence, down to a spreadsheet row or a table in a PDF.

That last detail is the connection back to the 95%. A citation only means something if the record it points to was integrated correctly in the first place, which is the part the notebook agents spend their hours on.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Peregrine counts its field engineers as R&D, not a cost center

An engineer who faked a missing editing feature using comment fields told Peregrine what to build next. A hurricane simulator stayed with one city. Co-founders Nick Noone and Ben Rudolph describe how they decide which piece of field improvisation becomes a product — and what they say it now costs to serve a city this way.

8 min read

Peregrine's pitch: make police data useful without owning it

On the Training Data podcast, Peregrine founders Nick Noone and Ben Rudolph argue that the public-safety software business has grown by collecting ever more data, and that their company inverts it: join the records an agency already holds, leave ownership with the agency, and lock down who may look. The same logic leads Noone to refuse a company-wide ban on facial recognition, leaving that decision to customers, law and local norms.

7 min read

Peregrine tested its first agent on a case detectives had already finished

On the Training Data podcast, Peregrine co-founder Ben Rudolph describes building the company's first operational AI agent with a police customer that had worked a case ending in the exoneration of a wrongly convicted man, then asked whether an agent could reproduce the same findings. He says the agent, which runs for 30 to 60 minutes over hundreds of gigabytes of case evidence, is now used in a few US departments, including a Wisconsin county where a handful of phone records helped place a suspect. Co-founder Nick Noone says the company deliberately lets customers take the credit.

4 min read

When one bad idea convinces all 5,000 agents

On Moonshots with Peter Diamandis, an operator described a bad coding idea spreading through a swarm of 5,000 identical AI agents until he intercepts and rewinds them — otherwise, he says, roughly $50,000 of tokens goes into a harebrained plan. The panel set that experience beside a new paper on "mind viruses" that spread between agents through editable memory, and argued about whether "virus" is the right word.

6 min read