AI Agents Daily Briefing: Sunday 25 October 2026
Good afternoon. Today's lead is a price war over coding agents. OpenAI is reportedly using deep discounts to pull big customers away from Anthropic, and one of them is moving its spending to Codex as a condition of the deal. Anthropic, meanwhile, has told companies building agents to stop stacking up subagents and fold them into one main loop. Two of the people behind Claude Code have left the company. And TechCrunch has rounded up the assistants that now live in your text messages. It's a quiet Sunday, so this is a short edition.
LEAD: OpenAI buys coding share from Anthropic with deep discounts
Background: Coding agents like Claude Code and OpenAI's Codex have been the biggest business in agents this year. Anthropic's models have been the default for many coding tools, which gave it pricing power.
What happened: The Information reported on 24 October that OpenAI is winning Anthropic's customers with steep discounts. Rippling's chief executive, Parker Conrad, said that in May Anthropic models handled 57% of coding work across his company's engineers and OpenAI handled 16%. Now OpenAI handles 52% and Anthropic 18%, a shift he put down to cost. Replit moved most of its AI spending to OpenAI in a deal worth more than $100 million a year, after OpenAI offered more than 30% off and Anthropic dropped the discount it had been giving. Training company DataCamp reportedly got 40% off on condition that it uses Codex rather than rival tools such as Claude Code or OpenCode, and its chief executive said costs could fall 80 to 90%. A US bank with more than $7 billion in assets had its ChatGPT seats doubled to 500 at no extra cost, paying under $7 a user instead of the usual $20 or more. An investor memo cited in the coverage says Anthropic's growth slowed from the second quarter to the third.
Why it matters: For many teams, coding agents are becoming interchangeable, and price is now deciding which one they use. Tying a discount to Codex shows OpenAI is competing for the agent tool itself, not just the model underneath. The figures come from anonymous sources and from customers describing their own purchases. Neither OpenAI nor Anthropic has confirmed them.
Sources: The Decoder · AIReport · The Information (paywalled)
Anthropic: fold your subagents back into one agent
Background: Many company agents are built as a "router" that hands work to a dozen or more specialist subagents. Anthropic's own earlier advice helped make that design popular.
What happened: Anthropic's Applied AI team has published new guidance with the opposite advice. Across several enterprise deployments, it says one main agent loop using "skills" frequently beat both a single giant prompt and the subagent design, often at lower cost and latency. One customer's runtime was roughly twice as fast. Another company's dozen specialist subagents made a consistent, on-brand customer experience hard and added latency. Anthropic warns against agents that copy the company org chart, against "Frankenprompts", and against "subagent sprawl", meaning hundreds of overlapping subagents. It says it cut Claude Code's own system prompt for advanced models by over 80%. It also says tool choice gets worse beyond about 30 to 50 tool definitions. Subagents still make sense when context bloat measurably hurts performance, for parallel jobs that need a fresh start, or when the agent struggles to pick the right tool. Anthropic is open about the risk. In its own test, folding a shipping subagent into a sample support agent cut latency but broke a process that had been working: the main agent answered from the order-lookup tool without checking the delivery date, and the failure showed up in only one of two runs. It has released new Claude API skills (build-eval, prompt-audit, hillclimb, cost-optimize) and a "Measure, Combine, Fold" migration kit on GitHub, labelled as an unmaintained reference implementation.
Why it matters: This changes the default way to design company agents: start with one agent and add helpers only when tests show you need them. The failure that appeared in one run out of two shows why Anthropic says to run evaluations before folding anything.
Sources: Anthropic · GitHub kit · The Decoder
Two Claude Code builders leave Anthropic
Background: Claude Code is Anthropic's coding agent and one of the best-known agent products. It runs in the terminal, on desktop and in the browser.
What happened: On Saturday, Thariq Shihipar posted: "I've left Anthropic after more than a year of working on Claude Code and seeing it scale larger than I ever imagined." He said he still believes in the mission, called his colleagues "the most talented, thoughtful, competent, and earnest people I have ever met," and said it was "time for a bit of a break before diving into something new." His post passed five million views. Later the same day Anthony Morris, who says he created Claude Code on desktop, wrote that Friday had been his last day. He called leaving "the hardest thing I've done in my career" and said he will "share what's next soon." He added that "there are still great folks working on the app."
Why it matters: These are two well-known Claude Code builders leaving in the same week Anthropic faces the price pressure in today's lead. Neither gave a reason beyond rest and new projects. Nothing links the two departures, Anthropic has not commented, and neither has said where he is going.
Sources: Thariq Shihipar on X · Anthony Morris on X
The agents that live in your text messages
Background: A growing group of personal assistants skip the app. You text them, and they carry out tasks for you.
What happened: TechCrunch published a guide on 24 October to the assistants that work over text messages. Instinct, which gives its assistants their own email addresses and is rolling out phone calls, raised $1 billion at a $10 billion valuation in September and is still in private beta. Wajo's Fo has its own email, phone number and payment card, and brings in a human when it gets stuck. Maverick asks permission before acting, and its Pro plan costs €27 a month. Proxy costs $1 a day. Poke, the first AI agent on Apple Messages for Business, was bought by Cognition in July for what was described as "low nine figures". Town raised a $55 million Series A in June.
Why it matters: Text messages are becoming a main way into personal agents, and investors are paying a lot for it. Giving agents their own phone numbers, inboxes and payment cards is exactly the kind of access the banks were worried about this week.
Sources: TechCrunch
Already covered in earlier briefings and not repeated here: the LiteLLM breach, Cursor Computer Use, Daytona's funding, banks blocking unvetted agents, 1Password for Agents, Claude Duet, Agent Skills moving to an open foundation, Replit's purchase of Chroma, General Reasoning's Avalon-M1, the Shandong harness study, and Claude auto mode becoming the default.