Urgent.News

What's breaking now, across thousands of outlets.

Tech

Pilotfish: Redesigning Browser Agents

An agent opens an invoice. The network is quiet. The DOM has stopped changing. The balance says $8,400 , under a banner that reads "Refreshing balances" . The wrong optimization is to return that number faster. The useful one is to cut repeated page text and model round trips without losing track of what the agent knows . The rule that makes the rest work: a diff must start from the page the…

An invoice appears before a network becomes idle. The displayed page's Document Object Model (DOM) no longer experiences changes. The account balance is listed as $8,400, situated beneath a banner labeled "Refreshing balances." The optimal approach is not to expedite that number; instead, it involves minimizing repeated page content and modeling round trips without losing track of the agent's knowledge.

The fundamental principle is that a diff must initiate from the page the caller originally received, rather than merely the page captured by the browser. This perspective is an excerpt from a technical guide authored by a Pilotfish team member, employing a hypothetical receivables portal. The customer, invoice, and amounts are representative.

The core components include maintaining row integrity, diffing against the page the model received, executing failure and recovery procedures, minimizing unnecessary typing, and testing specific scenarios. The published results demonstrate the effectiveness of these strategies. The guide emphasizes connecting to a browser for visual demonstration.

The source notes provide additional context about the plain-row path, which preserves cell order and reference action. The row serializer ensures cells are delineated, avoiding indiscriminate flattening of tables. A compact view may omit certain controls, necessitating the availability of a fuller read. The diff against the page the model received is crucial, as typing a customer name only modifies a field and suggests a few options.

Resending the unchanged invoice table is inefficient. The diff helps, only if the model can apply it, starting from a full observation that reached the caller. The base must be a complete observation, not the server's latest capture. Failure and recovery involve delivering a full observation as the base on delivery. If a response is withheld, a subsequent observation must be full again.

A diff also requires page order; swapping invoices should not appear as an unchanged list. When the document changes, the base becomes outdated, or the diff becomes too large or ambiguous, a full observation must be sent. Resynchronization is integral to the design. The failure and recovery process is tested using Node's built-in assertions.

The model is a simplified representation of Pilotfish's implementation, focusing on a small protocol model rather than an exhaustive integration test. Each balance represents a full page, and delivery is synchronous, requiring handling of dropped responses, acknowledgements, and resets.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

276,000 offers, 1 sale: vanity metrics vs real KPIs after 53 days

TL;DR I'm building Runometry , a price comparison site for sports gear, solo, with Angular/AnalogJS on Cloudflare. After 53 days: 276,000+ live offers, 800,000+ prices recorded, 2,800+ models.

  • Runometry generated 276,000+ offers and 800,000+ prices in 53 days
  • Only 900+ clicks led to 1 sale and €2.75 commission
  • Site struggled to index by Google despite 43,000 indexable pages

Get Uptime Kuma alerts as SMS from your own phone (no Twilio)

Uptime Kuma tells you when a site or a service goes down. Here's how to get that alert as a real SMS on your own number, sent by an Android phone you own: no SMS provider in between, no per-message…

  • Uptime Kuma notifies website/service downtime via SMS.
  • SMSTunnel app on Android phone enables text message alerts.
  • Free account provides 25 texts per day for monitoring alerts.

Tests that prove something: break a copy and see the test fail

A green suite only counts if it can go red. So when a test is written for a fix, the practice in Yonto Player's repo is to revert the fix and watch that test fail, and to break the exact line the test…

  • Tests verify bug fixes by reverting fixes and confirming test failures
  • Two claims: addressing behavior or establishing a new rule
  • Mutate code one decision at a time, revert changes to isolate impact

More from Thursday 19 November →