2026-08-24

Daily AI Digest: Slack gives coding agents team rooms; OpenClaw expands browser control

This is the GolemWorkers daily AI digest for people who want a practical view of what autonomous software can do now. Today’s edition covers agents collaborating with whole teams, operating browsers, coordinating recycling plants, mapping back-office work, preparing prescriptions for clinician review and testing vehicle software.

An operations lead coordinates autonomous work across an office, recycling facility, clinic and vehicle test bay.
A product team gathers around one shared project table to direct several coordinated building tasks.

Slack Code gives whole teams a shared room for coding agents

Slack Code lets a team summon Claude, Devin, GitHub Copilot, ChatGPT or Vercel agents into a dedicated code channel, where everyone can follow the work, review diffs and live previews, redirect the agent and approve high-stakes actions. Code channels are available on any Slack plan, but users still need access to the partner agent they choose.

Source: Salesforce
An operator supervises a secured relay connecting one paired browser station to several test devices.

OpenClaw’s new beta lets browser tools drive paired Chrome sessions

OpenClaw 2026.8.1-beta.3 adds a Puppeteer-compatible relay so external browser tools can control a paired Chrome session, plus verified SQLite backup and restore, safer gateway restart handoffs and GPT-5.6 model support. The release is a beta for people already running or evaluating OpenClaw rather than a stable-channel update.

Source: OpenClaw
A recycling plant operator watches coordinated conveyors and robotic sorters recover mixed materials.

EverestLabs launches agents that coordinate recycling plants in real time

EverestLabs Navigator combines vision systems and specialized agents to watch material streams, predict equipment trouble and coordinate sorters across recycling and recovery facilities. Plant operators can connect it to existing controls and legacy equipment, then use its recommendations and actions to reduce downtime and recover more plastics, fiber and metals.

Source: EverestLabs
An operations expert pauses at an ambiguous branch while a governed workflow waits for clarification.

Caddi watches back-office work, then builds governed agents for it

Caddi’s new agent identifies repetitive work in wealth management, law and insurance firms, ranks it by frequency and cost, then learns the real process while an experienced employee demonstrates it and explains exceptions. Each resulting agent follows auditable rules, connects to existing business apps and leaves every run replayable for oversight.

Source: Caddi
A physician reviews a prescription draft with a patient before giving final approval.

FDB turns exam-room conversation into prescriptions for clinician review

FDB Script Agent is now commercially deployed inside Tebra’s AI Note Assist for independent medical practices. It converts a patient visit into a structured, dose-aware prescription draft, but the clinician must modify or approve it before anything is sent, and unusual prescriptions require an explicit override.

Source: FDB
An automotive test engineer supervises an autonomous repeat test on a vehicle and physical control modules.

CANoe agents can write, run, repair and rerun vehicle tests

Vector’s CANoe 20 SP2 lets automotive teams describe a development or testing task in ordinary language, then assign agents to generate tests, run simulations, analyze failures, correct code and try again. Users bring the underlying model, monitor the process inside CANoe and choose how much autonomy the workflow receives.

Source: Automotive Testing Technology International

Agent idea of the day

Run a daily website change watch before customers notice

A website owner reviews before-and-after page evidence and prioritized repair cards from an autonomous monitor.

What this agent does

Catch broken pages, changed offers and unexpected signup or checkout steps before they cost your business trust or revenue.

Best for: Founders, operators and customer teams responsible for a public website, signup flow, booking path or online store.

Give it

  • Five to ten public URLs and the approved purpose of each page
  • A clean test account when login is necessary
  • The expected offer, price range and success condition for each journey
  • A named human reviewer and a strict no-purchase, no-publish boundary

Tell it to

  1. Open every approved page from a clean browser session and capture the visible state.
  2. Compare structure, offer details, links and the next required action with the last accepted run.
  3. Follow only the approved signup, booking or checkout path and stop before any irreversible submission.
  4. Separate meaningful changes from rotating content, personalization and harmless visual noise.
  5. Deliver evidence for each material difference and rank the three fixes most likely to affect customers.

Run it: Run every morning and immediately after pricing, authentication, checkout or campaign changes.

You get

A dated before-and-after evidence packet, broken-link and journey report, severity ranking and three recommended fixes for human approval.

Keep a human in control

  • Inspect only public pages and explicitly approved test accounts.
  • Never bypass CAPTCHAs, access controls, rate limits or terms of service.
  • Do not purchase, publish, message customers or change production data.
  • Redact personal data and secrets from screenshots and retain only the evidence needed for the report.

Feasibility: OpenClaw’s official documentation says its scheduler persists recurring jobs, wakes the agent at the required time and can deliver results to a chat channel or webhook. That supports the recurring inspection-and-report loop; this recipe limits the browser work to approved pages and keeps purchases, publishing and production changes behind a human decision. Source: OpenClaw Automations documentation →

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