Emmanuel will join our webinar on Ai-readiness on Thursday, Apr 23rd. Register to hear more of his insights here: https://lnkd.in/dHDSphpV
Your API, CLI, or even MCP is not the place to be unique. The most boring product wins for agents. I’ve been testing hundreds of developer tools with agents and have been surprised at what I found (full download link below): ▶ The more predictable, standards-based, and conventional your product is, the higher the likelihood of an agent being successful with it. ▶ Naming was always important, but with agents it’s even more so: if it’s their training data, they’ll understand it. ▶ A good error is better than good documentation. Agents try things all the time and can recover if they understand what’s wrong. ▶ Code beats anything else, including docs. Give an agent an SDK, sample code, a Skill that contains a snippet and it’ll use that. ▶ And the biggest surprise of all: Your docs are the most and least important thing at the same time. How’s that possible? Agents will either read part of your docs or sometimes none at all, if they think they have seen it before. So they may try to hit your API, for example, with obsolete knowledge from their training data. But if they do end up reading it, they may take it literally to the point of ignoring your schemas or other specs, so make sure the descriptions and workflows are accurate. So make sure your docs and specs are hyper-accurate - but don’t depend on an agent reading them. This is rapidly evolving and changing with every new model release and each new feature in MCP. I’m constantly updating what I teach in workshops, courses and consulting engagements. I put my current know-how for AI-ready APIs in a checklist that you can download below, but please give me feedback on your findings, so I can continue to make updates: → https://lnkd.in/ghCAR-wi PS - I have a similar checklist coming soon for MCP and CLI, so stay tuned.