2026-07-22

Daily AI Digest: OpenAI’s phone agents and Block’s shared workspace

This is the GolemWorkers daily AI digest, a concise look at practical agents that can finish useful work while people keep the final say. Today’s edition covers production voice support, shared human-agent workrooms, rehearsal coaches, security triage, factory robots, and a Codex release built for longer-running multi-agent projects.

An operations director supervises customer support, factory automation, security response, and team coordination from one bright control room.
A support specialist and supervisor review a complex caller case before a human escalation.

OpenAI Presence puts managed voice and chat agents on the front line

OpenAI Presence is now available in limited general availability for eligible enterprises that want agents to answer customers, verify accounts, use company systems, take approved actions and hand sensitive cases to people. OpenAI says the system already resolves 75% of calls to its English-language support line without a human handoff, while each customer sets permissions, tests and escalation rules.

Source: OpenAI
A small cross-functional team coordinates shared work around a project board and several task stations.

Block opens a shared workspace where people and agents work side by side

Block released Buzz, a free open-source workspace with channels, threads, voice, media, code repositories and automated workflows. Teams can bring agents powered by Codex, Claude Code, goose or other systems; each agent gets its own portable identity and defined permissions to post, review code, run approved automations and contribute alongside people.

Source: Block
An employee practices a difficult conversation with a remote roleplay coach while a manager observes.

Synthesia’s AI coach now rehearses difficult workplace conversations

Synthesia launched Roleplay Sessions for enterprise training, letting employees practice sales calls, performance reviews and customer complaints with an interactive avatar that talks back and pushes back. After each attempt, an AI coach scores the conversation against a company rubric, explains what worked and tracks progress for managers.

Source: Synthesia
Two security analysts hand off an unlabeled evidence packet in a controlled night operations center.

Swimlane’s AI SOC triages security alerts across many client environments

Swimlane launched an AI security operations center for managed security providers. Its agents normalize, enrich and correlate alerts into an explainable verdict and investigation plan, while a central workbench keeps client data separated and records the approvals, evidence and human decisions needed before response actions move forward.

Source: Swimlane
A production engineer supervises a safe component handoff between two wheeled industrial robots.

Humanoid funds coordinated factory robots for late-2026 deployments

Humanoid raised

52 million to begin beta deployments of its wheeled industrial robots in the fourth quarter and prepare mass manufacturing. Its KinetIQ system lets multiple robots communicate and hand off related tasks, and manufacturing partner Schaeffler has agreed to deploy 1,000 to 2,000 units across its facilities by 2032.

Source: Humanoid
Three engineers pass an audio recorder between separate, color-coded task stations.

Codex 0.145 turns voice notes and old sessions into coordinated agent work

Codex 0.145 adds audio input and stabilizes its opt-in multi-agent experience with configurable models, reasoning levels, concurrency and restored roles. Developers can also import settings, MCP servers, plugins, sessions, commands and project memories from Cursor and Claude Code instead of rebuilding their working context by hand.

Source: Codex release notes

Agent idea of the day

Build a morning opportunity scout for contracts and grants

A small-business owner and colleague verify municipal bid notices, deadlines and a regional project map.

What this agent does

Monitor the public sources that matter to your business, verify new opportunities and deliver a short evidence-backed list before deadlines become urgent—without contacting buyers or submitting anything.

Best for: Small contractors, agencies, consultants and nonprofits that cannot spend hours checking scattered bid, grant and procurement portals.

Give it

  • A fixed list of official portals, buyer pages and grant sources to monitor
  • Your service area, capabilities, minimum and maximum project size, required certifications and no-go sectors
  • A calendar of blackout dates plus the person who decides whether an opportunity moves forward

Tell it to

  1. Check only the approved source list and record every new or changed notice with its canonical URL.
  2. Extract the buyer, scope, location, value, eligibility rules, required documents, question deadline and submission deadline.
  3. Score each opportunity against the provided geography, capabilities, capacity, certifications and deal-size limits, citing the evidence behind every score.
  4. Reopen the official notice before reporting it, flag amendments or missing documents and discard expired, duplicated or aggregator-only listings.
  5. Prepare a ranked brief with the five strongest fits, the first three actions for each and a clear approval request; never contact a buyer or submit a response.

Run it: Run at 7:00 a.m. every weekday and immediately when an approved portal posts an amendment to a tracked opportunity.

You get

A one-page opportunity brief with verified links, fit scores, deadlines, missing requirements, recommended next steps and an explicit pursue-or-pass queue for a person.

Keep a human in control

  • Treat the official notice and its attachments as the source of truth; never rely on an aggregator for dates or eligibility.
  • Never contact buyers, register an account, make promises, set pricing or submit a bid or application without explicit approval.
  • Do not invent certifications, past performance, team capacity or financial information; mark unknowns visibly.
  • Escalate legal terms, bonding, insurance, conflicts, unusual data requests and any deadline ambiguity to a person.

Feasibility: Cascade’s launched service monitors state, city, district, county, federal and private construction sources, connects prior tender data and recommends projects a firm is positioned to win. That live workflow supports a narrower, safer agent that monitors an approved source list, verifies canonical notices, scores fit and prepares a human decision brief without outreach or submission authority. Source: TechCrunch coverage of Cascade →

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