AI News — Morning — 2026-10-22

Omnara moves its agents off the laptop

• Omnara introduces a self-hostable service for long-running assistants — Its new platform separates the running assistant from the machines it controls, so closing a laptop need not stop the work.
  • An agent is a model running a multi-step task with tools. Omnara’s October 21 introduction describes a new platform replacing its earlier remote-control app for coding assistants.
  • The Apache-2.0 project stores events in PostgreSQL and runs the model loop on interchangeable workers. Waiting agents release their workers until more work arrives.
  • A daemon on each connected Mac or Linux machine makes outbound connections. The agent can use several machines, including a laptop and a cloud sandbox.
  • Tools can run immediately, require human approval or be blocked. The platform supports your own model endpoints, including Ollama, alongside hosted providers.
  • Self-hosting requires four long-lived services, a migration job, PostgreSQL, Redis, object storage and a shared memory filesystem. Docker Compose packages the local setup.
  • Caveat: The local defaults are intentionally insecure, and operating the platform yourself includes securing its services and connected machines. (HN 82 points · 25 comments)

Agent access & purchasing

• Walmart gives approved shopping assistants a dedicated way in — A supported interface and separate spending permissions offer a more controllable integration than automating an ordinary shopping session.
  • Agent Gateway uses MCP, the standard for connecting agents to tools, for inventory searches and purchases by approved agents.
  • Agent Profiles are scheduled to roll out next month. Each links an assistant to the shopper’s account and separates its purchases from the shopper’s own transactions.
  • Customers can require approval for each transaction or set a spending limit. They can delete a profile, and Walmart+ membership is unnecessary.
  • Walmart says its normal returns process applies when an agent acts unexpectedly. Its terms prohibit agents from accessing payment information or exceeding the shopper’s intent.
  • Caveat: This announcement covers approved agents, with the customer-profile controls still awaiting their November rollout. (Techmeme)

Tests that check the checker

• AI-written task checkers often reject perfectly valid solutions paper notes → — VerifierBench shows why a checker needs tests of alternative correct answers, alongside tests that it catches mistakes.
  • A benchmark is a fixed set of tasks for comparing models. This October 19 paper contains 222 tasks and 4,993 hidden cases spanning program requests, resulting application states and code changes.
  • Across 19 models using the same surrounding software, generated checkers accepted 21.1–70.2% of valid solutions. They rejected 79.2–96.3% of invalid solutions.
  • Checkers that rejected every tested solution caused 87.9–96.0% of false rejections in the analyzed archived records. A checker can look strict while failing its basic purpose.
  • Six audit checks cover replay, repeatability, source tracking, consistent labels, trivial baselines and deliberately modified checks. Source defects removed 45 of the original 267 tasks.
  • Caveat: The detailed failure analysis uses selected traces and archived records, so it identifies mechanisms without measuring their prevalence across every configuration. (arXiv 2610.28679)

Memory: cost and trust boundaries

• Searching raw history can cost less than building memories paper notes → — MemoryTax gives memory-system builders a small, reproducible comparison that includes the often-hidden cost of constructing the store.

  • The October 21 study tests seven methods with three models on one conversation containing 419 turns and 152 questions. Each combination runs once.
  • With Claude Opus 5.5, Pi scored 99% for $6.95, versus 95% for $17.84 when reading the full history. Costs include model and text-indexing requests, excluding grading.
  • Pi also received photo captions withheld from the memory methods. On the 88 questions without photo evidence, Pi scored 100% and full context scored 99%.
  • Building Mem0’s store with Opus alone cost $25.88. The author’s projected break-even against full context requires 596 questions on that same history.
  • Caveat: One conversation that fits every model’s context window cannot establish that searching raw history will replace memory systems generally. (arXiv 2610.29823)

• Malicious saved instructions can keep affecting fresh chatbot sessions paper notes → — The study separates attacks through ordinary personal memory from attacks requiring control of shared projects or organization settings.

  • The October 20 paper reports April–June tests of ChatGPT Instant 5.3, Claude Sonnet 4.6 and Gemini 3 Flash. Saved instructions caused persistent wrong answers or simulated unavailability.
  • Researchers planted a document in a shared Claude project that created an unwanted memory for another project user. Editing shared project instructions or knowledge requires the corresponding project permissions.
  • Logging coworkers’ prompts to an attacker-controlled Google Drive required owner-level access to establish the connector. Organization instructions then directed the logging.
  • An attack counted as successful when its goal failed in all five pre-injection attempts and succeeded in any of five afterward. Three author-reviewers voted on outcomes.
  • Caveat: These are earlier product versions and permission-dependent demonstrations, rather than proof that any outside document can compromise today’s chatbots. (arXiv 2610.28810)

Document processing

• Dropping document colors makes assistants miss the author’s warnings paper notes → — Testing extracted words alone can miss information lost when a converter strips highlighting and color from policy documents.
  • On 76 documents, the authors’ formatting-aware representation scored 0.690 on recovering marked requirements. Docling scored 0.150 and plain text scored 0.075.
  • Those values are micro F1, which balances finding the marked requirements against listing the wrong ones across the collection. Formatting-preserving docx2python scored 0.586.
  • Removing one visual cue at a time identified color and highlighting as the main contributors. Bold, italic, monospace and strikethrough each changed the measured score by at most 0.003.
  • The parser recovers cues without model requests, but its representation uses roughly twice the tokens of plain text. Assigning meanings to cues helped less consistently across models than preserving them.
  • Caveat: The task specifically asks for author-marked requirements, and its markup-derived labels still await human validation. (arXiv 2610.30026)

Model access

• OpenAI plans to sell models through Google Cloud — Organizations with Google Cloud spending commitments will gain another purchasing route for OpenAI models and Codex workflows.
  • The October 21 announcement places access in Model Garden, with purchases transacted through Google Cloud Marketplace and eligible for existing cloud commitments.
  • OpenAI also plans availability in Google Antigravity and Gemini Agent. Its announcement describes general availability in the coming weeks and offers an access-request form.
  • The stated integration pairs model access with Google’s access controls and data governance. The announcement leaves exact model pricing and regional availability unspecified.
  • Caveat: This is a planned rollout, so it supplies neither immediate general access nor verified terms for moving sensitive workloads today. (HN 294 points · 122 comments)