You might have noticed these landing from me rather than William over the past few weeks. He's taking a break, so these will be coming from me going forward. If anything here sparks a thought or you just want to chat, hit reply. I'd love to hear from you.

Welcome to this week’s edition of Twimbit X Insights.

Four frontier vendors spent five months arriving at the same conclusion about where the value in an AI agent actually sits.

Four vendors, one conclusion

Picture a team nine months into building the machinery that keeps an agent running: the retries, the memory that survives a long task, the traffic control when several agents work at once.

That machinery is now something you can buy. Anthropic opened Claude Managed Agents in public beta in April. AWS previewed its managed harness for Bedrock AgentCore the same month and made it generally available in June. Google followed at I/O in May, and on 10 September OpenAI put its Agents API into public beta, opening the engine behind Codex and ChatGPT Work to any developer.

Five months, four vendors, one shared conclusion.

Which leaves a better question: what is left that nobody will sell you?

What comes in the box

OpenAI runs the engine and charges nothing for it, billing only the tokens, tools, and compute an agent uses. You still choose where the agent works, which tools it can reach, and what data those tools open onto.

The early results show how capable the bought layer has become. SafetyKit reports a 60% reduction in cost per case after shifting a review workflow onto it. Hypha says moving to the managed engine cut failed agent responses by 86%. Both are company-reported.

Results that looked like a moat in March are now available to your competitors, and to every vendor booking time in your calendar.

What stays on your side

Think about the AI work already on your roadmap for next year.

Plenty of teams will keep building their own runtime, and often for good reasons: data residency, an existing cloud commitment, control the managed versions do not offer yet. Either route is defensible, and neither is where your advantage comes from.

Two agents running on the same engine still answer differently, and the gap sits in what each was given to work with. An agent that has read three years of your deals reaches conclusions a general model never will.

That part also survives the churn. The engine will be replaced within a year, but your pricing history, your win and loss patterns, and the reason a Jakarta deal behaves nothing like the same deal in Singapore carry across every switch, improving quarter by quarter while the engine improves on someone else's schedule.

Renting the plumbing buys speed. Owning the context decides whether that speed is worth anything.

Which one did yours fund?