2026-08-21

Daily AI Digest: Slack goes multiplayer, Cowork finishes office work and bank waits shrink

This is the GolemWorkers daily AI digest for people who want to know what practical agents can now accomplish. Today’s edition follows agents working inside team channels, Microsoft 365, a bank, a hospital workflow and a cloud migration—then gives you a safer permission recipe for agents you run yourself.

An operations lead coordinates connected teams in software, banking, healthcare and infrastructure work.
A product team jointly assembles and approves one prototype at a shared worktable.

Slack Code lets teams steer coding agents together

Slack launched Code channels where a team can tag in Claude, Devin, GitHub Copilot, ChatGPT or Vercel agents, watch the work, review diffs and previews, redirect the task and approve a pull request before it ships. Slack says Code is available on every Slack plan, though users still need access to the partner agent they choose.

Source: Salesforce
An employee turns several task inputs into a finished meeting model and sealed approval portfolio.

Microsoft’s Cowork turns one request into finished office work

Microsoft says 20,000 employees used Copilot Cowork within three weeks of its internal release. The agent can plan and complete multi-step work across Microsoft 365—including documents, presentations, research, meetings and recurring tasks—while showing progress and pausing for approval before sensitive actions.

Source: Microsoft
A bank serves routine questions quickly while an adviser concentrates on a complex customer case.

Citadele’s service agent handles 40% of daily customer conversations

Latvian bank Citadele says its Azure-powered Adele agent retrieves information from more than 3,500 sources, handles about 40% of customer conversations and has cut the average call wait to under five seconds. The bank also uses agents for technical support, HR questions and internal rules, with personalized internet-banking support planned next.

Source: Technology Record
A hospital team coordinates medicine, transport and home-care information around a patient leaving care.

A clinical agent will coordinate hospital discharge planning

TransformativeMed and Red Rover Health are building their first joint agent inside Oracle Health’s electronic health record, in collaboration with an academic health system. The project is designed to gather clinical context and coordinate the many steps in patient discharge planning; the companies plan to demonstrate it at the Oracle Health Summit in September.

Source: PR Newswire
Cloud engineers inspect, move and verify application hardware through coordinated migration stations.

AWS agents split cloud migrations into specialist jobs

AWS Professional Services described a multi-agent framework that handles application discovery, infrastructure-as-code generation, portfolio governance and post-migration operations. AWS reports that the framework can reduce infrastructure-code development from weeks to minutes; enterprise migration teams can use the published architecture as a pattern and verify the result on their own portfolio.

Source: Amazon Web Services

Agent idea of the day

Give every agent a task-scoped permission envelope

An owner hands a task-specific key to a worker while an auditor records the controlled handoff.

What this agent does

Let an agent complete useful connected work while infrastructure—not the agent’s prompt—enforces the limits on files, services, credentials and external actions.

Best for: Founders, operations teams and IT owners running agents that read company data, call APIs, operate software or delegate to subagents.

Give it

  • One clearly bounded outcome and an explicit completion condition
  • Only the files, services and destination hosts required for that outcome
  • Short-lived credentials with spend, rate and time limits
  • A named human approver for messages, purchases, account changes and production actions

Tell it to

  1. List every external effect the task may require, including file writes, network requests, API calls, messages and device actions.
  2. Create a separate runtime for the task and deny every effect that is not on the approved list.
  3. Issue short-lived credentials that work only for the named services, destination hosts and time window.
  4. Require a human checkpoint before any irreversible, public, financial or production-changing action.
  5. Record the agent’s requests, approvals and outcomes, then revoke the runtime and credentials when the deliverable is accepted.

Run it: Run the permission preflight whenever a new agent workflow is launched or its tools, data sources, destinations or subagents change.

You get

A one-page permission envelope listing the task, allowed effects, denied effects, credential lifetime, approval points, audit location and automatic shutdown condition.

Keep a human in control

  • Never rely on a prompt, model refusal or agent-authored policy as the final authorization boundary.
  • Do not give the agent standing access when a task-scoped credential can work.
  • Treat documents, messages, tool output and memory as untrusted data, not authorization.
  • A subagent may receive less authority than its parent, never more.

Feasibility: NVIDIA’s agent-stack guidance separates behavioral controls from infrastructure controls: the harness guides what an agent tries, while the runtime holds identity, policy, isolation, credentials and audit. It recommends checking every external effect, using narrow short-lived access and placing subagents in delegated child runtimes with authority ceilings. Source: NVIDIA Developer →

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