The agents that do the work.
OpenAI shipped ChatGPT Work and Anthropic pulled Claude chat and Cowork into one tab, both reframing their products around agents that do the work, not chatbots that answer. For a fund, the edge is the layer you build and own around the model.
Hi folks,
This week both Anthropic and OpenAI changed their products' entry points. OpenAI shipped ChatGPT Work, Anthropic pulled Claude chat and Cowork into one tab, and both now open onto AI that actually does work.
The one thing
The agent that ships finished work, not a draft
- OpenAI launched ChatGPT Work, an agent (running on GPT-5.6) that:
- Takes an outcome, pulls context across your apps and files (Slack, Teams, Drive, SharePoint, Salesforce, CRMs, calendars),
- Works for hours breaking a goal into steps,
- Asks approval before sensitive actions,
- Returns finished work, memos, spreadsheets, decks and reports, and runs scheduled tasks (OpenAI).
- It looks a lot like Anthropic's Cowork, but simpler: OpenAI's is one interface built for the job, whereas Cowork sits inside the wider Claude app.
- This week Anthropic moved the same way, pulling Claude chat and Cowork into a single home tab (TechCrunch).
- Both point one direction: the labs are showing what their frontier intelligence actually does: get work done, rather than answer questions like a classic chatbot.
What this means for a fund:
- The unit of work shifts from "a draft I finish" to "a deliverable I review": first-cut IC memos, portfolio updates, LP-reporting drafts, data-room summaries.
- We already see this in the work we are doing with funds.
- The approval gate and per-connector permissions give a compliance-minded fund a governance surface to reason about.
- The output is only ever as good as the context and checks you wrap around it (the reason allowing connectors is vital).
In the mix
- BlackRock is steering private credit into AI-infrastructure lending (Bloomberg, PE Wire)
- BII head Jean Boivin called AI capex a "big tailwind" for private credit on 7 July; Bloomberg Intelligence puts the six largest US hyperscalers' 2026 capex near $820bn, up almost 80% year on year.
- The top models are open to everyone again (OpenAI)
- GPT-5.6 went GA for everyone on 9 July after its government-gated preview.
- As noted, Fable was back the week before too.
- Palantir's Karp called the labs' model broken and pushed "AI sovereignty" (CNBC)
- On CNBC, Alex Karp said "something has gone completely wrong": enterprises burn escalating token costs for little value and hand their IP to the labs.
- Palantir put out a 9-point AI-sovereignty manifesto, and some firms are moving to open-weight models at a fraction of the cost.
- "Own your data and your stack" is moving from fringe to boardroom.
- For a fund it reframes AI vendor risk as a sovereignty and data-ownership question.
From my week
- I spent the second half of the week in San Francisco, thinking a lot about self-improving agents.
- The model (and to an extent intelligence) is becoming a commodity: everyone can buy the same one. We believe the durable edge is the layer around it, the compounding context, the checks, and the playbook, that you build and own.
- This essay by Lilian Weng calls this the "harness" and argues that near-term progress comes from improving that layer, not the model. The loop is:
- Find where it failed,
- Make a bounded fix,
- Validate it,
- and treat the context as a living playbook rather than a prompt written once.
- That is exactly what turns a generic agent into one that ships reliable fund work.
- It is also the layer we build inside the fund and own, with the governance, auditability, and data residency, baked in from the start rather than bolted on.
Levercon is the AI team for investment funds. We embed alongside your team to automate workflows, build custom tools, and compound operational gains. Ultimately we help Australian based investment funds leverage AI.
Book a strategy call.
See where AI fits in your fund.
We'll show you where AI lifts efficiency and performance across your organisation.