The Agent Stack #004 — Wednesday Stack

Anthropic just shipped Claude Cowork plugins for finance, engineering, and design work. This isn’t another AI assistant announcement. It’s the first serious attempt to replace actual SaaS workflows with agents.

I’ve been testing the Google Workspace integration for three days. The promise is simple: tell Claude to “analyse Q4 expenses and create a budget proposal”, and it connects to Sheets, pulls data, runs calculations, and drafts documents. In practice, it’s more like having a very capable intern who needs constant supervision.

The technical execution is solid. Authentication works smoothly. The agent handles multi-step workflows without breaking. I watched it pull financial data from three different spreadsheets, cross-reference against a project management doc, and generate a formatted budget proposal in under two minutes.

But here’s what the demos don’t show: edge cases kill everything. When my expense sheet had merged cells, Claude got confused and started hallucinating numbers. When I asked it to “format this professionally”, it created a presentation so generic it looked like clipart from 2003. The agent needs perfect data and crystal-clear instructions to work reliably.

The Docusign plugin tells the same story. It can create contracts from templates and send them for signature. But it can’t handle custom clauses or negotiate terms. It’s automation, not intelligence.

Still, this matters because Anthropic is targeting workflows that cost enterprises millions. Financial reporting, contract management, design reviews. If Claude can handle even 60% of these tasks reliably, that’s transformational for operating costs.

The competitive landscape shifted overnight. Salesforce, Microsoft, and ServiceNow all have agent platforms in beta. But Anthropic’s approach feels more practical. Instead of building a new interface, they’re plugging into tools people already use daily.

OpenAI’s COO admitted this week that “we have not yet really seen AI penetrate enterprise business processes.” Anthropic is betting they can change that by starting with boring, repetitive tasks rather than trying to replace entire job functions.

Quick Hits

• MatX raised £380M to challenge Nvidia with chips designed specifically for transformer inference. Founded by former Google TPU engineers, shipping silicon in 2026.

• Meta signed a £76B AMD chip deal with warrant structure, diversifying beyond Nvidia while chasing “personal superintelligence” goals.

• Nimble raised £36M for AI-powered web scraping that validates and structures real-time data for agent consumption. Could solve the stale training data problem.

One Thing to Try

Set up Claude Cowork with your Google Workspace account. Start with one simple workflow—like generating weekly status reports from project data. Document every failure mode. The pattern of where it breaks will tell you exactly where human-AI collaboration works in your organisation.

The boring enterprise stuff might just be where agents finally prove their worth.