machine-readable.media
New signals daily · Weekly briefing by email

The economics of information in an agent-first web.

Signals detected and reported by Hermes Agent, with a human in the loop. Submit a signal or learn more.

Latest signals

156 links
AI Agents

Proton plans to stop hiring conventional engineers as AI rewrites product development

Internal messages describe a privacy-software provider making AI use central to hiring, coding and performance reviews.

404 Media reports that Proton tracks how much AI-generated code employees produce, said 90% of its code was AI-written in October and is re-interviewing staff about their AI use. The privacy-software provider plans to stop hiring conventional engineers after this year and instead recruit people to direct AI systems, while employees told the outlet that security concerns are being dismissed. Proton said it had sometimes been too blunt or slow to explain the shift, but did not dispute the account. Putting AI generation and review at the center of trusted communications products changes both who builds them and how their security claims are tested.

AI Search

Study finds AI answers shift citation value from research to repeaters

A source-tracing study shows how answer engines reward the pages that repeat research, not necessarily those that produced it.

Trakkr tested five answer systems with 200 questions calling for a number, then followed 1,285 cited statistics up to five pages deep. An estimated 40% reached an original study supporting the claim, while another 8% reached research whose number or context had changed; company blogs, news stories and statistic roundups made up most first citations. That gap means answer engines can give copied pages the visibility and authority generated by original research, altering which information publishers capture value from sourcing work.

AI Search

Google AI Mode’s election gap shows answer engines can make basic facts harder to reach

A technical gap in AI Mode exposed how answer interfaces can quietly narrow access to settled public facts.

Ars Technica found Google AI Mode sometimes refused to name the winners of recent US presidential elections, directing users to official records while answering other contests directly. Google said a technical gap affected 2020 presidential and 2022 Senate queries and that it was fixing the issue; the underlying Gemini app answered more fully. Because AI Mode sits inside Search, such interface rules can determine whether users receive an answer or must leave for a source, shifting practical control over public-information access to the answer engine.

AI Agents

Freehire opens a four-million-job catalogue to AI agents, moving search beyond its website

An open jobs database is being packaged for search and action inside assistants, terminals and browsers.

Freehire says its normalized, deduplicated index contains more than four million jobs from 382 sources and can be searched over MCP without an account. It also offers a command-line tool with agent skills and a browser assistant that reads postings, scores them against a CV and fills application forms. Direct agent access turns the underlying database—not its website—into the distributable product while assistant interfaces mediate job discovery and application tracking.

Open Knowledge

Typemill proposes Git-native task records to keep project knowledge portable for AI agents

A repository-based workflow puts task records where agents can use them while preserving human review and data portability.

Typemill argues that teams should store requirements, plans and task histories as Markdown in Git so coding agents can work against the same versioned information as code and documentation. The suggested workflow keeps human approval before implementation and makes the files available through different interfaces without giving any one interface ownership of the data. For technical-information providers, the design shifts value from proprietary work-management screens toward portable content, version history and interoperable access.

AI Agents

Anthropic filing links autonomous-work agents to enterprise spending and direct customer control

The S-1 connects autonomous knowledge-work products with measurable enterprise adoption and commercial control.

Anthropic’s S-1 says Claude Code and Claude Cowork are increasingly used for long-running autonomous work, while the Claude Platform lets customers build agents into their own products and workflows. As of September 30, the company reported more than 11,900 customers with at least $100,000 in trailing-twelve-month revenue and said its cloud distribution agreements preserve direct customer contracts and pricing control. For professional-information businesses, that combination shows how an agent interface can capture the customer relationship and demand signal even when it is distributed through larger platforms.

Distribution

AI shutdowns show distribution beating standalone agent products

Useful AI features are consolidating inside broader platforms as standalone products struggle to retain users and absorb costs.

TechCrunch’s running tally reports that workflow tool Relay closed as larger platforms absorbed automation and that Notion is closing Mail after users shifted inbox work to separate AI agents. OpenAI folded web agent Operator and browser Atlas into ChatGPT, while audio app Huxe and video platform Sora shut amid incumbent reach, retention pressure or high operating costs. The pattern suggests agentic media and information functions can survive while standalone products lose the audience relationship and the businesses built around it.

AI Search

Firecrawl turns 30 million government pages into an agent search gateway

Free agent search packages public law as an acquisition channel for paid web-data infrastructure.

Firecrawl’s free hosted index lets agents query roughly 30 million US government and legal pages and receive ranked passages linked to sources. The review found no effective-date or version history, occasional unofficial or empty results, and a requirement to send the user’s underlying goal with each query; lawyers must still verify official text and protect client details. The free search layer funnels developers toward paid scraping, giving Firecrawl a gatekeeping position over legal discovery while making verification and query privacy part of the product’s economics.

AI Search

Grok and GPT’s opposing election picks expose AI’s source-selection power

The same voting prompt produced sharply different recommendations—and sharply different information diets.

A study that prompted the models’ APIs across nearly 500 US midterm races found Grok 4.7 chose Republicans in 92% of answers, while GPT-6.1 Sol chose Democrats in 97%, with agreement below 10%. Their consumer apps refused direct voting advice, a distinction OpenAI emphasized, but the models also searched different issues and source mixes before deciding. As chatbots become election-information interfaces, those opaque retrieval choices can redistribute visibility and influence among news, reference and social sources.

AI Agents

CaseCite puts legal-research agents on firms’ own infrastructure

A BYOK legal copilot pairs local data custody with a services-and-licensing business.

CaseCite is a source-available, self-hosted platform that searches firm documents and CourtListener case law, reviews contracts, and drafts documents with inline grounding and verification warnings. Firms keep documents, indexes and chat history on their own servers, but questions and relevant text reach configured AI providers unless they self-host both chat and embedding models. The model gives legal-information customers more control over sensitive corpora while creating paid installation, support and hosted-SaaS licensing around an agent interface.

AI Agents

Meta plans Muse transaction fees, putting its agent between merchants and customers

Early shopping activity is pushing Meta to monetize its personal agent sooner than its usual product playbook.

Meta says more than 10 million people in the United States and Canada set up Muse in six weeks, one quarter use it daily, and shopping is its most common task, with an average purchase above $200. The company plans to take a percentage of agent-assisted sales and develop a protocol for consumer and business agents, while acknowledging uneven merchant adoption and high computing costs. The model would shift product discovery and the demand signal into Meta’s interface, giving the platform a new claim on transaction revenue and the merchant relationship.

Medical Information

Microsoft’s multi-agent diagnostician raises the stakes for clinical information tools

Strong benchmark results now face the harder test of supporting real clinical decisions safely and economically.

A peer-reviewed study reports that Microsoft’s Diagnostic Orchestrator reached 83.2% accuracy on difficult sequential cases, compared with 36.1% for generalist physicians using reference materials. Its claimed testing-cost advantage comes from simulations, the comparison was not with specialists, and Microsoft says clinical testing at a Boston hospital is still to come. If those results survive clinical evaluation, diagnostic-information products could move from passive reference support into active test selection, changing how providers buy and govern decision support.