2026-07-23

Daily AI Digest: Trulioo traces owners, Procore deploys 20 agents

This is the GolemWorkers daily AI digest, covering practical agents that investigate, coordinate and complete real work while people retain control. Today’s edition follows business-ownership research, construction workflows, regulated customer service, high-volume finance operations, computer-agent safety and a durable new agent runtime.

An operations lead reviews an evidence folder at a shared studio that connects construction, customer service and secure document work.
Two compliance analysts trace a cross-border chain of company records to one accountable owner.

Trulioo’s agent reconstructs who really owns a business

Trulioo launched its UBO Discovery Agent for compliance teams that need to identify the people behind complex companies when registries stop at another corporation or stale filing. It rebuilds cross-border ownership chains from governed sources, cites every finding and is designed to abstain when it lacks a defensible answer; Trulioo says a large social-commerce customer began deploying it across Malaysia, Vietnam and the U.K. this month.

Source: Trulioo
A superintendent, scheduler and safety lead review plans and site photos beside an active construction project.

Procore packages 20 agents for construction teams

Procore made three Digital Coworker packages generally available, ranging from five starter agents to a 20-agent library that reviews safety images, schedules, change orders, contracts, bids and project records. Administrators get usage controls, company-specific Skills start rolling out in August, and early adopter Haskell told Procore that one submittal-review workflow fell from seven days to 10 minutes.

Source: Procore
A member-services specialist helps a customer complete a coverage change with a supervisor available for escalation.

Ushur’s agents carry regulated customer requests through to completion

Ushur announced a platform for insurers, health plans and financial institutions to build agents that gather information, retrieve documents, act across company systems and complete customer journeys across text, email, web, chat and voice. Initial uses include health-plan member service, Medicaid redetermination and transportation requests, with auditability, human oversight and more than 200 integrations built into the system; access begins through an early-access waitlist.

Source: Ushur
A financial operations team routes one complex case from an organized work queue to a senior reviewer.

Yiren Digital puts numbers on a large-scale agent rollout

Yiren Digital says its agents now handle eligible asset-recovery, voice, marketing and other high-volume workflows through a shared multi-agent platform spanning six business functions. In the company’s reported asset-recovery results, manual handling fell from 45.0% to 24.9% and tickets handled per staff member rose from 358 to 525, while human handling remains part of the workflow.

Source: Yiren Digital
A cable slips from a sealed glass workroom toward keys stored in a drawer outside the approved boundary.

Researchers find a path from Claude Cowork’s local VM to Mac files

Accomplish AI demonstrated a vulnerability chain that let a local Claude Cowork session leave its Linux virtual machine and read or write files outside the folder a Mac user had connected. The latest Cowork uses cloud execution by default, where the researchers say this path does not appear to apply, but the local design still makes the approved-folder boundary an important safety check for computer-use agents.

Source: Accomplish AI
A technician assembles planning, memory, approvals and observation instruments into one durable modular field case.

Microsoft releases a ready-made runtime for long-running agents

Microsoft released its Agent Framework Harness for Python and .NET, bundling tool calls, plan-and-execute state, per-model-call history persistence, context compaction, file memory, skills, web search, approvals and OpenTelemetry into one configurable runtime. Background agents, direct file tools, looping and shell access remain opt-in preview features rather than part of the stable core.

Source: Microsoft

Agent idea of the day

Build an offer desk that turns your expertise into sellable services

Two independent professionals compare three carefully prepared service packages against portfolio examples in a sunlit studio.

What this agent does

Interview you about work you already do well, turn the evidence into a few tightly scoped service offers, research realistic price ranges and prepare an approval-ready listing packet without publishing or promising anything.

Best for: Independent consultants, experienced operators, freelancers and people testing a side business who have useful expertise but no clear service menu.

Give it

  • Your résumé, portfolio, past project summaries and any testimonials you have permission to use
  • The kind of buyer you want to help, the outcomes you can reliably deliver and the work you refuse
  • Your available hours, preferred turnaround, geographic limits and minimum acceptable fee

Tell it to

  1. Interview me about repeatable results, difficult problems I have solved and the evidence I can show for each claim.
  2. Extract three service ideas with one target buyer, one concrete outcome and a narrow scope for each.
  3. Draft what is included, what is excluded, the inputs required from the buyer, the delivery format and a realistic turnaround.
  4. Research comparable public offers and cite the sources behind a recommended price range; never invent demand or set a final price.
  5. Prepare a listing draft, buyer intake questions, risk flags and an approval table; do not publish, message prospects or accept work.

Run it: Run when you add a strong portfolio example, change your availability or review the service menu each quarter.

You get

A three-offer decision pack with evidence-backed positioning, scope boundaries, public pricing comparisons, intake questions and a final approve-revise-reject queue.

Keep a human in control

  • Use only work history and testimonials you are allowed to share; redact client names and confidential details by default.
  • Do not claim credentials, results, availability or experience that the source material does not support.
  • Keep pricing advisory and let a person approve the final fee, terms, taxes and refund policy.
  • Never publish a listing, contact a buyer, accept a job or promise a deadline without explicit approval.

Feasibility: SpringBrand’s live marketplace launched an AI Co-Founder that interviews sellers, helps package skills into fixed-price services, provides pricing guidance, drafts listings, handles initial buyer questions and clarifies scope while users approve each step. That public workflow supports a narrower offer-preparation agent that stops before publication, outreach or commitment. Source: SpringBrand AI Co-Founder launch →

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