- Sarah Simionescu of Composio said many startups made their sites friendly to AI search, but few made their applications usable by agents.
- She argued that an agent arrives with a goal and judges a product on one thing: whether it can get the job done.
Sarah Simionescu, a member of technical staff at Composio, argued that many startups are preparing for the wrong visitor. On the AI Engineer show, she said many have tuned their landing pages and sites for AI search, known as GEO. Few, she said, have made their products ready for AI agents to operate.
What was said
Simionescu's talk was about the death of the dashboard. She said people never wanted a dashboard or its query language. They wanted the answer. She traced the path from five tools with five query languages, to AI features added as sparkle buttons, to MCP, a way for agents to connect to apps.
In her view, MCP alone still falls short. She named three faults. Agents do not pick up lessons over time. A long tool list overwhelms the model. Each app stands apart from the others.
In a live demo, an agent ran an analysis across PostHog and Metabase. It never loaded the full results into its context window, which is the amount a model can hold in mind at once.
Her lesson came at the end. Composio, she said, now hears from startups whose own customers are pressing for a way to use their services through agents. She described the new audience as "a new species of user." It arrives with a goal and a set of tools. In her words, "it judges you on exactly one thing, whether it can get the job done."
Why it matters
Our reading: AI search work and agent work are two different tests. A site can be easy to find and easy to quote, yet hard for an agent sent to finish a task. Her demos, which moved from a bug report posted in Slack to a proposed code fix, show the kind of task she means.
For teams that build software, the work sits in the application itself. It covers what actions an agent can take and how much data it must haul around to take them. For teams that buy software, a useful question follows. Can our own agent finish a real task in this product, or only read its marketing page?
The other side
Simionescu works at Composio, which turns apps into tools that agents can use. She has a stake in the argument. Her claim that "so few" companies have prepared their applications was an impression, not a figure. The requests from startups were also reported secondhand, with no count given.
She did not say GEO work is wasted. That conclusion comes from the framing, not from her. The talk also leaves open how widely agents will be sent to use products, and how a company would measure being easy for agents. Her answer, as a demo-led pitch, was the interface Composio is building.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Dashboards Are Dead — Sarah Simionescu, Composio (2026-10-04)
