AI Engineer World's Fair 2025
Just do it. (let your tools think for themselves) - Robert Chandler
Loading slides…
Read the talk
Just do it. (let your tools think for themselves)
A failed Slack message exposes the cost of tiny API tools. Robert Chandler shows how a specialized workflow can give a general agent a simpler, more useful interface.
From a talk by Robert Chandler
Before you start: Familiarity with LLM tool calling and MCP, the Model Context Protocol, will help you follow the workflow design.
Why is sending Philip a message so hard?
Sending a Slack message to Philip should be a straightforward task for an agent. Robert Chandler introduces the problem from his experience as Wordware’s co-founder and CTO: he says he has helped hundreds of teams build reliable agents, after previously working on self-driving cars. The promise is familiar—an automated system understands you and your team, interacts with your tools, and puts useful output where you need it.
His early experience connecting Slack through MCP fell well short of that promise. He wanted to tell Philip that he was using MCP and thought it was cool. The agent listed users, became confused, listed channels, and attempted to send a message. Even after reportedly finding Philip, it resorted to posting in the general channel, asking someone else to pass the message along. Chandler reports that this unsuccessful attempt to message Philip took about five minutes. The failure points to the central design problem: many MCP servers expose low-level API operations that were never designed for a language model to coordinate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Small tools make the caller do the integration
A large JSON response can be convenient for a deterministic program: code selects the fields it needs and moves on. For an agent, that same response becomes more material in the context window. When tools each expose only a small function, the model must also work out how to compose them, interpret every intermediate result, and decide what to call next. The integration logic becomes part of the agent’s reasoning burden.
Pagination makes the problem concrete. One response may not contain the user or record being sought, so the agent must request another page and continue until it finds the desired data. Each iteration extends the chain of requests and adds more output to the context. Authentication adds a separate kind of overhead: hosted MCP servers had made setup somewhat easier, but Chandler still encountered API keys and Slack bot creation. His expectation that this friction would diminish over the following months was a forecast at the time of the recording.
Tool definitions themselves also consume attention. Chandler’s historical examples are eight tools from a Slack MCP and approximately 20 more from a Notion MCP. Every additional tool brings instructions and another choice for the agent; even those two integrations do not cover the full range of useful automation. The issue is both the number of available operations and the length of the sequences needed to accomplish a task.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make the tool a specialist
Chandler illustrates the mismatch with a T-Rex holding a tiny spanner, then Inspector Gadget carrying a thousand tools. A powerful caller can still struggle when its tools are too small or too numerous. His alternative is a team of Avengers: Hulk handles smashing, Hawkeye handles precision, and Iron Man might serve as the broadly capable main agent. The useful part of the analogy is specialization—each delegate owns a meaningful task.
An agent can be a tool for another agent. The outer agent needs a tidy natural-language interface and a useful result; the specialist can own the reasoning and sequence of operations behind that interface.
| Design choice | Function-sized tool | Specialist tool |
|---|---|---|
| Scope | One operation | A meaningful task |
| Coordination | Caller assembles the sequence | Delegate manages the work |
| Result | Intermediate API output | Task-level output |
The intended payoff is a reusable interface that produces reliable, high-quality results without requiring the caller to reconstruct the workflow each time.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Package competitor analysis as one task
Wordware MCP Toolbox supplies the implementation in the demonstration: it turns Wordware workflows into tools that other agents can call. Chandler selects a competitor-analysis workflow from the landing page because the task combines judgment, reasoning, and integration with both Twitter and Notion.
The workflow uses Wordware’s built-in Twitter scraping tool, avoiding a separate search for a Twitter MCP server. Its analysis instructions specify what Chandler cares about instead of leaving the entire definition of a good report to the model. He could add more company context to sharpen the comparison and identify where the businesses differ. This is where the specialist acquires its task-specific judgment: the workflow encodes the criteria before the general agent invokes it.
The workflow then generates the analysis, writes it to Notion, and returns the created page’s URL. Its contract includes both producing the content and delivering it to its destination. In the editor, the completed sequence shows analysis guidelines followed by Generate Analysis Report and Create Page, with an instruction to output the Notion page URL. The calling agent can receive that link without managing each publishing step itself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect the toolbox and request the analysis
The demonstrated Toolbox service is in beta; Chandler describes a broader rollout as coming soon. He uses an existing toolbox containing the competitor-analysis app he had already published, then connects it to Claude. A toolbox can group several tools, whether related or disparate, and different toolboxes can be switched on or off for different tasks. That gives the user a way to control which capabilities are available to the caller.
The invocation is a short natural-language request:
Create a competitor analysis for Anthropic AI.
Claude selects the Wordware tool and presents permission choices to allow it once or always. Chandler judges the task low risk and authorizes it to proceed. The workflow is now responsible for carrying out the analysis, rather than asking Claude to assemble Twitter retrieval, analysis, and Notion publishing from individual operations.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Return a result the user can find again
Chandler then introduces a prepared result: “Here’s one I made earlier.” He follows its Notion link to a formatted competitor summary, which he describes as based on all of Anthropic’s tweets. The page illustrates the intended output; it does not establish that the newly requested run has completed or independently confirm the retrieval coverage.
The report offers observations about Anthropic’s tweeting style and what someone could learn from it. Just as useful is its location: the analysis is already formatted and stored in Notion, where Chandler wants to retrieve it later, instead of being lost in a chat. The destination is part of the task’s definition of success.
Chandler characterizes the resulting tool as highly reliable, repeatable, and aligned with the task, though he supplies no quantitative evaluation. The architectural benefit is clear: a generic agent gains a specific capability by delegating to a workflow whose instructions and integrations have already been assembled for that job.
This returns to the team analogy. A generalist delegates work to a specialist with the context and capabilities to carry it through. Wordware Toolbox is one way to build that boundary, but Chandler explicitly leaves the implementation open: any platform can support agentic MCP tools. The design invitation is to give the tool enough responsibility—and enough room to reason—to complete a useful piece of work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Setup page for exposing Wordware workflows to MCP clients such as Claude Desktop. Requires a Wordware API key.
Further reading
- MCP tools specification — March 2025 revisionDocumentation
The contemporary protocol contract for tool discovery, invocation, schemas, results and user confirmation.
- Archived Slack MCP reference serverRepository
Historical eight-tool Slack implementation documenting pagination, channel messaging and bot setup. The repository is archived.
Updates since the talk
Notion's server repository documents API changes and directs users toward its remote OAuth-based MCP service.
The founders describe Wordware's origins and its later shift from workflow infrastructure toward the Sauna assistant.
Full transcript
Loading transcript…