Subscribe Forward this edition

The Agentic Enterprise.

AK · Daily Edition · 8 min read

Friday, October 23, 2026

Claude wants to work before you ask.

A proactive assistant changes who decides a task is worth starting and how often the human gets pulled back in.

Claude Duet, announced by Anthropic on October 21, brings a proactive assistant into a waitlisted research preview for US Pro and Max subscribers. It puts the choice of what deserves attention into the product itself. For leaders watching where enterprise work is heading, that raises a practical question: how much judgment should an assistant exercise before it asks for yours? The current preview offers an early place to examine that boundary.

The Big Story

Proactive agents

Claude wants to work before you ask.

C

laude Duet can prepare a morning brief and check connected accounts every few hours, returning when work needs attention. Its cloud computer can keep working while the user's devices are off. Anthropic says the system is designed to show a proposed action and wait for approval before sending messages, paying, signing or deleting. Those are product claims, not independently verified safeguards. Anthropic announcement

Consider an unpaid invoice. An assistant could find it, prepare a reminder and bring the draft back for approval. The useful judgment begins earlier than the send button: is this the right account to chase today, and does the proposed reminder reflect the customer relationship? An approval screen helps only if the reviewer has enough context to answer.

The interruption is part of the product.

That makes the quality of the queue worth examining. A busy account owner may dismiss an alert because it arrives at the wrong moment. A team evaluating this kind of assistant should record which interruptions produced useful decisions and which merely added another inbox to manage.

Today's limits matter: Duet isn't available on Team or Enterprise plans; expansion is planned without a date. Microsoft 365 is unavailable, and users can't watch or take over its browser. Periodic checks can also miss urgent arrivals. These constraints make the preview a limited test of the working relationship, with broader enterprise deployment still ahead. Getting-started guide

The Spearhead Take

Start with a contained queue where a person can check both the recommendation and its supporting record. Keep responsibility for urgent exceptions explicit while you learn whether the assistant's interruptions earn their place in the working day.

Sources: Anthropic announcement · Getting-started guide

The Obvious & The Overlooked

What is easy to see, and what is easy to miss.

The Obvious

Background work consumes a budget: Duet's shared plan limits make unattended activity part of a user's capacity planning before the next requested task begins. Source: Anthropic support

Robotics buying includes operations: Formic's recurring model calls for a named budget owner who covers ongoing service after the installation team leaves. Source: Formic

Shared definitions need business agreement: Sakana's ontology approach makes settling what the company's data means part of selecting an agent platform. Source: Sakana AI

The Overlooked

Global HR templates need local variants: Italy's final-decision rule makes jurisdiction-aware workflow configuration a practical requirement for multinational employers. Source: PwC

Revocation belongs in the acceptance test: Glean's per-user connector deserves a test in which a worker loses source access during a workflow. Source: Glean

Disclosure needs ongoing ownership: The UK's planned police-AI register will need someone to maintain entries as deployed tools change after the first release. Source: Home Office

Moving Pieces

Five developments with practical consequences.

Identity

The agent can log in without holding the password

1Password launched credential access for personal agents on October 22: the user approves a login on their own device and credentials are filled directly into the destination, according to the company. Current availability covers Individual and Family accounts on 1Password.com; several integrations, including ChatGPT Work and Claude Duet, are still forthcoming. Keeping a password away from a model is useful, but buyers must separately check which actions a logged-in agent can perform and whether their account is supported. Treat each integration's approval behavior as a separate acceptance test. 1Password

Employment

Italy reserves the final decision for a person

Italy's Decree 179 takes effect today, October 23, and Article 41 requires a person with effective autonomous judgment to make final employment decisions. Automated sourcing and shortlisting are excluded from that final-decision definition; employee consent doesn't waive the restriction. For an employer using AI in Italian operations, a review button alone won't establish that a person exercised judgment, so the implementation needs to preserve the basis for the decision and the human's ability to change it. Document the explanation route before the first disputed decision arrives. Official Article 41, PwC analysis

Security operations

A generated fix still waits for a merge

AWS made Continuum for code vulnerabilities generally available on October 22 in Northern Virginia and Oregon, using code and AWS environment context to assess exploitability and propose pull requests. Every fix currently requires human approval and merging; autonomous remediation remains a future direction. This can change the queue a security team works through, but the production outcome still depends on someone reviewing the patch, running the relevant checks and confirming the vulnerable path is closed after deployment. Assign the merge queue to someone whose workload allows timely review. AWS

Physical AI

Factory automation includes the service call

Formic announced $50 million in growth financing on October 22, combining Series B equity and debt, to expand a robotics business that bundles equipment with deployment and ongoing service under recurring payments. Expansion toward general-purpose robots is planned, rather than evidence that those systems are already widely operating. The useful buying question is how responsibility works when a line stops: a recurring contract should make maintenance response and production acceptance concrete before the customer depends on the equipment. Specify the response obligation in hours your plant can actually use. Formic, Axios

Enterprise deployment

Uber gives the rollout a support structure

Uber began expanding ChatGPT Work on October 22 from a 1,000-person pilot toward access for more than 30,000 employees, with an internal skills library, approved integrations and dedicated training. The planned adoption and outcome reports remain ahead, so the workforce figure describes intended access rather than proven use or productivity. A deployment team can borrow the operating question now: who maintains a shared skill when employees report that it fails? Give the workflow's owner a clear route for those reports before extending access. OpenAI, TechCrunch

On the Radar

Nine signals worth a glance.

Enterprise knowledge

Sakana AI announced Aquarium on October 23 around shared business definitions and governed execution; test how changed definitions affect workflows, with pricing and GA timing unpublished. Sakana AI

Inference

You.com launched Inference Router on October 22 with model serving and web search in one API; assess hosting choices and contractual terms together. Company announcement

Knowledge

Glean made its ClickUp connector generally available on October 22 with per-user authorization; access checks still belong in the deployment review. Release notes

Accountability

Fifth Circuit vacated an injunction and ordered case reassignment on October 22 after fabricated material and other errors; verification must reach the factual record. Opinion

Administration

GitHub began enforcing Copilot global defaults for eligible unconfigured features on October 22, through October 26; explicit settings remain preserved. Changelog

Public sector

UK Home Office updated police-AI guidance on October 22 with a public-register pilot intended by year-end; public access hasn't launched yet. Government guidance

Research operations

Warp Sciences emerged on October 22 to build automated chemistry services; the plan puts physical experiment turnaround into the AI discovery discussion. Company announcement

Biotech

Carbon-X released downloadable DNA-and-language model weights on October 22; the model card's clinical-use exclusions matter for teams evaluating local research tools. Model card

Creative operations

Canva released Craft in Dream Lab and its public API on October 22; editable outputs belong in a brand-review workflow before reuse. Canva

Quick Hits

The wider field, one line each.

Workflows

01Cube added a Claude connector for permissioned financial models and transaction-level traceability on October 22. Announcement
02WebMD Ignite announced Ruby on October 22 for connected healthcare-provider marketing workflows using approved data. Announcement
03Cognizant announced AI Nexus on October 22, describing an operating layer for AI cost and quality management. Announcement

Research and logistics

04Genedata added NVIDIA's RFD4-Proteina model to connected protein-design and laboratory-result workflows on October 22. Announcement
05Körber announced a stake in NAiSE on October 22 to expand orchestration of mixed warehouse-vehicle fleets. Announcement
06OpenAI and NASK signed an agreement to explore AI safety and security research collaboration on October 23. Announcement

Visibility

07Trust3 AI released its Claude Enterprise API integration October 22; the Claude Team telemetry mechanism is forthcoming. Announcement
08GitHub added names for internal skills and agents to Copilot usage metrics on October 22. Changelog
09Mave announced financing on October 22 to expand agentic security operations across customers' existing tools. Announcement
10CoreWeave announced customer key and support-access controls October 21; confidential-computing expansion remains planned. Announcement

The Number

15%

The share of surveyed Goldman Sachs 10,000 Small Businesses graduates who said AI was fully integrated into core operations.

Babson collected 2,473 alumni responses in August; the report warns they may not represent all graduates, so this single study shouldn't become a national adoption estimate. Report

Counter-Signal

Industrial readiness

Enthusiasm doesn't staff the factory

Hitachi's October 22 physical-AI report surveyed more than 200 global industrial decision-makers. While 97% reported some use, just 8% placed their operational deployments at continuous optimization and innovation. These are vendor-survey self-assessments, with detailed fieldwork methods unavailable in the public summary. Hitachi

A team can reasonably support AI and still lack the people needed to keep an installation useful. Before accepting an ambitious factory roadmap, ask the operating manager what happens during a bad shift, including who diagnoses a failed handoff and how the next shift learns about it. Budgeting for that work is a practical test of readiness. A demonstration can't answer it for you.

From the Field

An assistant has to earn the interruption.

I'd begin with a queue whose consequences the team already understands. Unpaid invoices are a useful example because an account owner can judge whether the proposed follow-up is timely and whether the underlying record is right. Keep the existing collection process available while you examine the assistant's choices.

Then look closely at what comes back for review. A short recommendation should include the record that supports it and enough context to let the owner change course. If people keep opening several other systems before they can decide, the approval task still has work left in it. Ask the reviewer what would make the next decision easier.

I'd also discuss the quiet periods. Someone must still notice an urgent customer message that arrives between checks, and that person needs to know what the assistant has already done. Set a small review period and compare the useful interruptions with the dismissed ones. Let those results shape the next workflow you try.

Attention is an operating resource too.

Let's create value with AI,
AK

The Agentic Enterprise

Know more about AI than 95% of your peers. Every morning.

Published by Spearhead. We build AI systems that work.

Spearhead partners covered include Anthropic, OpenAI, AWS, Microsoft via GitHub, NVIDIA and Glean. Produced with Codex at AK's request; draft awaiting AK's approval.

Spearhead · Contact · Subscribe

© 2026 Spearhead. All rights reserved.