2026-08-23

Daily AI Digest: Research agents reproduce papers and trading agents gain controlled accounts

This is the GolemWorkers daily AI digest for people who want to know what autonomous software can actually do now. Today’s edition follows agents reproducing scientific work, trading within user-set boundaries, testing live websites, building governed data pipelines and coordinating coding tasks.

A research lead reviews evidence returned by several autonomous work teams.
A scientist compares a reproduced experiment with the original paper while an autonomous research team works nearby.

Faraday independently reproduces published research

Inherent introduced Faraday, an AI research teammate designed to reproduce findings from scientific papers without being shown the answer. The company says the agent, built around a 27-billion-parameter model and using Codex for coding work, outperformed larger Anthropic and OpenAI systems on its replication test; access is currently aimed at research teams.

Source: TechCrunch
A human account owner watches an autonomous market operator work inside a clearly bounded trading booth.

Binance lets AI agents trade through controlled subaccounts

Binance launched Agent OS so ChatGPT, Claude Code, Codex and other compatible agents can analyze markets, inspect account data and place trades. Users assign an agent to a subaccount, choose its permissions and decide whether every order needs approval; withdrawals are blocked by default, but Binance does not impose a separate trading-loss cap inside the subaccount.

Source: TechCrunch
Specialist work crews transform raw records through controlled stages while a data lead approves the finished pipeline.

AWS publishes a multi-agent path from raw data to governed pipelines

AWS published the Agentic Data Operations Platform, an open reference architecture that has one onboarding agent dispatch specialists for metadata, quality checks, transformations and orchestration. Data teams describe a new source and their controls, review the generated artifacts, then promote deterministic code into production instead of keeping a model in the runtime path.

Source: Amazon Web Services
Six distinct task couriers attempt the same customer journey while an evaluator records where each one stops.

Ora sends six major agents through live websites to find where they fail

Ora runs Claude Code, ChatGPT, Gemini, Hermes, OpenClaw and Vercel’s eve through real signup, integration and payment journeys, recording the cost, latency and exact step where each agent stalls. Companies can use those traces to repair workflows that human visitors may complete easily but autonomous customers cannot.

Source: Vercel
A technical lead coordinates several isolated software tasks from one physical operations board.

Codex adds one dashboard for starting, naming and stopping agent tasks

Codex 0.149 adds an interactive agents dashboard for searching, starting, opening, renaming and stopping tasks, plus a queue command that sends new instructions to existing local or remote sessions. The release also restores permission profiles on resumed and forked work and improves routing for sub-agent approvals and notifications.

Source: OpenAI Codex

Agent idea of the day

Build a weekly agent mystery shopper for your website

A website owner studies a route map returned by an autonomous mystery shopper.

What this agent does

Test whether an autonomous customer can discover, understand and complete your most important website journey before a real prospect gets stuck.

Best for: Founders, product teams and online businesses that rely on signup, booking, quote, checkout or support flows.

Give it

  • One public journey with a clear success condition
  • A test account and non-production payment method when needed
  • The pages and actions the agent may use
  • A maximum runtime, request budget and named human reviewer

Tell it to

  1. Start from a clean browser session and attempt the journey exactly as a first-time visitor would.
  2. Record every page, decision, action, delay and error without bypassing access controls or site rules.
  3. Stop before any irreversible purchase, submission or external message unless the human reviewer approved that step.
  4. Classify each failure as missing information, unclear interface, blocked automation, broken integration or policy boundary.
  5. Return the shortest reproducible path, screenshots, elapsed time, estimated cost and one prioritized repair list.

Run it: Run weekly and after any material change to signup, pricing, checkout, booking or authentication.

You get

A dated journey replay, failure map, screenshots, cost and latency summary, and the three highest-value fixes for human approval.

Keep a human in control

  • Use test data and approved accounts; never submit real personal, financial or health information.
  • Do not evade CAPTCHAs, rate limits, access controls or terms of service.
  • Do not make purchases, publish content, contact customers or change production data without explicit approval.
  • Keep credentials scoped to the single test and revoke them when the run ends.

Feasibility: Vercel reports that Ora already runs six agent harnesses through live signup, integration and payment journeys while tracing cost, latency and the exact point of failure. That demonstrates the core loop; this recipe narrows it to one approved journey, a clean browser, reversible test data and a human-controlled stopping point. Source: Vercel →

GolemWorkers — OpenClaw ready to work: launch in seconds, no setup, online 24/7.