Khe Hy says Codex's redesigned app leapfrogged Claude Code for many users because it is easier to use, faster and cheaper.
Khe Hy argues Codex running on a local machine reachable from a phone turns an always-on Mac mini into a host for personal agents, which is why OpenClaw users are switching.
Brett Caughran is skeptical of overnight agent demos, warning that 24-hour runs on due diligence produce long-running slop that burns tokens and feeds investor laziness.
Investors Move From AI Answers to AI Work Systems
AI for investors is moving from faster answers to systems that do work, but generic agents on generic data only produce slop. Value comes from personalised skills, proprietary data connected through MCP and research unified across scattered tools. Brett Caughran expects a mid summer inflection as stronger models and cheaper open source economics meet a maturing connectivity ecosystem.
Takeaways
Codex's local, phone-reachable agents reshuffled the tools race. Khe Hy says OpenAI's redesigned Codex app beat Claude Code for many users on ease, speed and cost. It runs locally and is reachable by phone, so an always-on Mac mini hosts agents. Firms add Codex rather than drop Claude.
Cheap analysis lets firms do work they never did. Khe Hy says training has moved from use cases to what is possible. In private equity, agents could scrape KPIs from hundreds of confidential memorandums. An associate's time costs about $100, while AI at about 90% quality costs about $20.
Unsupervised agent runs produce slop, not insight. Brett Caughran doubts overnight multi-agent demos, since a 24-hour due diligence run just burns tokens. He warns that summarising 10-Ks or generating theses from notes feeds laziness. Newer models, he says, are work mechanisms rather than answer engines.
Investors want different things, so personalisation wins. One fund buys on management quality while another trades revenue guidance revisions. Brett Caughran favours dashboards and custom skills over one product. His 100 page primer gave way to twenty-sixty-twenty, and he is more bullish on the application layer.
Proprietary data through MCP is essential for serious work. Generic chatbots pull primers from Zacks, Motley Fool and Seeking Alpha. Brett Caughran calls the FactSet MCP essential, though FactSet fell almost 50% during the Claude episode. Third Bridge and GLG are MCP native, and research integration is a commercial problem.
Scattered research needs headless systems built from the bottom. Hedge fund research sits in OneNote, Slack, Bloomberg IB, Outlook and PDFs, and the Microsoft 365 connector cannot read spreadsheets beyond 16 columns. Tinkerers build markdown skills first, then knowledge graph vendors package the workflows.
Five layers of a good investing system
Triage of incoming information.
Execution of the analysis itself.
Decision support, a headless system that analyses prior deals and operating memorandums.
Rigor enforcement on the investor's own process.
Idea tracking over time.
Tap a name to go there · tap a line to see why they are linked