Topic
Tool design
Naming, argument shape and descriptions that a model can actually use.
16 posts
- Does your agent work eight times out of eight?Average accuracy is the wrong number for a product. Run the same task eight times and count how often it worked every single time.Safety11 min
- Prompt injection when the agent can spend moneyIn a consumer app the untrusted text is your own user's data and the tools move their money, which makes filtering useless and structure the only real defence.Safety10 min
- Server-driven UI was already the answer. LLMs made it urgent.Server-driven UI and generative UI are one idea arriving from two directions, and a decade of mobile practice already tells you which parts are hard.Render spec10 min
- What an in-app AI agent actually is, and what it can touchThree different things get called an AI agent in a mobile app. Here is the one that lives inside your product and acts through your own backend.Agent basics11 min
- Where the seconds go in an in-app agent turnWe measured a slow turn end to end. Most of the time was not in our code and not in the model, which changes what is worth optimising.Tools and MCP10 min
- Letting users reorder by voice in three tool callsRepeat purchase is the cheapest first agent feature to ship. Three tools carry it, only one of them writes, and a confirm card sits between the second and the third.Agent basics6 min
- Agent or chatbot: telling the two apart before you buyA support bot answers questions. An in-app agent finishes the task. Five questions that separate them in any vendor demo, and where a chatbot still wins.Agent basics10 min
- Connect an AI agent to the backend you already runYou do not need a new API surface for an agent. You need to describe the one you have, and decide what it is never allowed to touch.Tools and MCP11 min
- MCP for in-app agents: connecting your tools to a modelThe Model Context Protocol replaces one connector per backend with one protocol. What it is, how a tool call actually happens, and what it costs on the first turn.Tools and MCP10 min
- Designing a tool surface an agent can actually useTwo independent vendors put the practical ceiling near twenty tools. Everything else about tool design follows from the model reading your schema and nothing else.Tools and MCP9 min
- In-app search when the query is a whole sentenceHow search inside an app breaks when people speak in sentences, and how to build the search tool an agent can actually call.Render spec7 min
- Agents in a travel app: rebooking is the feature worth buildingBooking is already a form. The task an agent is uniquely good at is rebooking under disruption, and it needs four read tools before it can answer anything.Business8 min
- The architecture of an in-app agent, for people who approve itEvery box in the system, who owns it, where the trust boundaries sit, and what changed between the intent-classification era and the one where a model chooses.Tools and MCP9 min
- How far you actually get in a day with an in-app agentA working spoken turn takes an afternoon. This is what those hours buy, what week two costs, and why the integration is not on the critical path at all.Integration9 min
- Agents in a shopping app: reorder, track, returnThe three retail tasks worth giving an agent, why reorder is first, and how to check whether your catalogue supports the case at all.Agent basics7 min
- Why pre-LLM assistants failed, and what changedAssistants before 2023 broke at the layer that turned words into actions. Three things replaced it: open intent, named tool calls, and a loop that checks first.Agent basics10 min