{ "video": { "id": "YiFqcu9YA38", "title": "Dashboards Are Dead — Sarah Simionescu, Composio", "duration": 642, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "YiFqcu9YA38", "title": "Dashboards Are Dead — Sarah Simionescu, Composio", "one_liner": "Sarah Simionescu, whose team builds the Composio dashboard, argues that dashboards and their query languages were only ever translators between people and their data. Agents now do that translating, but native MCP servers fail them in three ways: agents don't learn, too many tools drown the context, and every app is isolated. So products should be designed for agents, 'a new species of user' that has no eyes and judges you only on whether it can get the job done.", "summary": "The talk opens with a confession. After using Datadog every day for six months, Simionescu opened its dashboard for the first time and froze, because she had only ever reached it through an agent. She gives a 'postmortem' of the dashboard: in 2022, one bug meant five tools, five UIs and a different query language for each. In 2023 came the sparkle button. In November 2024 came Anthropic's MCP. Native MCP setups are still a mess, because agents don't learn, thousands of tool definitions drown the context, and each server knows only about itself. Through mockup demos of Claude using Composio's MCP (Slack to Sentry/Datadog to a PR, and a PostHog-to-Metabase analysis), she argues that an interface built specifically for agents is what fixes this.", "key_points": [ "2022 workflow for one Slack bug report: five windows and tools (Slack, Datadog, PostHog, VS Code, GitHub), each with its own UI to relearn after every redesign. Each also has its own query language: Datadog syntax, Jira's JQL, Slack search modifiers. Even tools that 'speak SQL' don't agree on SQL.", "Every dashboard and obscure query language was 'a translation device between you and your data', because the machines couldn't understand what you wanted. 'You wanted the answer.'", "2023's 'sparkle button' had an LLM write a 'sometimes correct' query, as long as the question needed no more than two database joins.", "MCP (announced by Anthropic in November 2024) is only a channel. How a service chooses to talk to agents 'makes all the difference'. With a dozen MCP servers wired up, three problems show: (1) agents don't learn and every conversation starts from zero, with skills as a mere band-aid; (2) more tools and skills mean more context, which makes agents dumber, and Composio's GitHub toolkit alone has 200+ tools; (3) every app is isolated, so tasks that span two apps are left to the user.", "Demo 1 (mockup): she pastes a Slack message link into Claude with Composio's MCP and asks it to use Sentry and Datadog, find the root cause and draft a PR. Composio search states three tasks, and for each one Composio returns the right tools plus a plan for using them (e.g. find the Slack channel ID before querying messages). Claude pulls Datadog and Sentry in parallel, scans the codebase, and has a PR up in under five minutes, with no custom workflow or skill.", "Early, unreleased results compare Composio with each app's native MCP from the Claude marketplace, using the same tasks and the same model, and show 'a clear difference' (no numbers given).", "Demo 2: Claude runs a PostHog query for the onboarding vertical distribution, then cross-references the e-commerce users' IDs with Metabase to find their most-used toolkits. It saves the IDs without loading them into context, samples the Metabase schema, and uses Composio Remote Workbench to generate an SQL query with a regex string, all in a few minutes.", "Startups make landing pages AI-friendly for GEO, but few have prepared their applications to be used by agents. Composio now gets requests from startups whose clients are 'begging' to use their services through agents." ], "takeaways": [ "Design your product's interface for agents, not just humans. The agent judges you only on whether it can get the job done, and it won't click your sparkle button.", "Don't assume shipping an MCP server is enough. Curate the tools rather than dumping hundreds of definitions into the agent's context, and give it plans that include dependencies (e.g. resolve the channel ID first).", "Keep large intermediate results out of the context window. Save them and process them in a remote sandbox or workbench, for example by generating SQL that matches the saved IDs.", "Tackle cross-app tasks with a unified layer rather than isolated per-app MCP servers, so the agent can join data across sources such as PostHog and Metabase." ], "topics": [ "agents", "mcp", "tool-calling", "context-management", "developer-experience", "observability", "agent-native-design" ], "tools": [ "Datadog", "Composio", "Composio Remote Workbench", "Slack", "PostHog", "VS Code", "GitHub", "Jira (JQL)", "Sentry", "Metabase", "Claude", "Anthropic", "MCP (Model Context Protocol)" ], "quotes": [ { "text": "You never wanted a dashboard or it's cursed query language. You wanted the answer.", "at": "02:27", "url": "https://www.youtube.com/watch?v=YiFqcu9YA38&t=147s" }, { "text": "MCP gave agents a door into every app. But it left them standing in thousands of separate rooms with no map and no memory of ever being there.", "at": "04:45", "url": "https://www.youtube.com/watch?v=YiFqcu9YA38&t=285s" }, { "text": "You are now serving a new species of user. They don't have eyes. They are not going to click your sparkle button.", "at": "09:44", "url": "https://www.youtube.com/watch?v=YiFqcu9YA38&t=584s" }, { "text": "For a decade, we've built pro uh we've built products to be easy for humans to use. And the next era belongs to those that are easy for agents to use.", "at": "09:59", "url": "https://www.youtube.com/watch?v=YiFqcu9YA38&t=599s" } ], "words": 2036 }, "summary_url": "/#YiFqcu9YA38", "transcript": { "html": "/transcripts/YiFqcu9YA38.html", "txt": "/transcripts/YiFqcu9YA38.txt", "vtt": "/transcripts/YiFqcu9YA38.vtt" } }