Agent teams win by moving complexity out of the agent
Vercel, Manus and Pi got better results by removing tools from the agent. The orchestration belongs in deterministic factory steps around a small, focused agent.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 14 April 2026 Updated: 1 min read
Everyone is adding tools to their AI agents. The winners are removing them.
Vercel’s text-to-SQL agent went from 80% to 100% success on its five benchmark queries. Not by upgrading the model, but by removing 80% of their specialized tools.
Removing features beat adding them
Manus (acquired by Meta) rebuilt their agent framework five times. The biggest gains every time came from removing features from the agent, not adding them.
Pi Agent runs on 4 core tools and a 200-token system prompt. That’s it.
The agent stays small. The factory does the rest.
Where the complexity goes
The point is less “less is more” than where you put the complexity.
Inside the agent, you get multi-step orchestration, tool routing and elaborate prompt chains. Every step is non-deterministic. Every tool competes for context window space.
Outside the agent, you get deterministic scripts that combine small, focused agent steps, with quality gates between the steps. The factory orchestrates. The agent executes one task well.
This is the Dark Factory principle applied to agent design. The agent is one station on the line, not the entire line.
What most teams get wrong
- 40-tool MCP configurations that consume half the context window
- Multi-step orchestration chains inside the agent that break when one step hallucinates
- Custom wrappers around every API call
- Elaborate prompt templates that fight the model’s native capabilities
What works
- Minimal agent harnesses with specific, focused tasks
- Deterministic scripting between agent steps: code, not prompts
- Quality gates at every station, not guardrails stuffed into the agent
- Factory orchestration that scales with compute, not with bigger context windows
The harness matters more than the model
Your MCP list is probably too big. Your agent is probably doing work that belongs in the factory. The harness matters more than the model. And the best harness is the one that does exactly one thing well, inside a factory that does the combination.
Are you stuffing orchestration into your agent? Or building the factory around it?
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Coding agent setup
- A harness is a list of things the model cannot do yet A harness is not plumbing around a model. Every component in it encodes one assumption: a thing the model cannot be trusted to do alone. Read it that way and the harness becomes a checklist.
- Claude Code can now navigate your code like an IDE Claude Code used to search a codebase with grep and text matching. With Language Server Protocol plugins it gets real code intelligence: references, signatures and types.
- The best agent architecture is a simple while loop Claude Code and the pi agent show the same thing: a main loop, a few general-purpose tools and careful context management beat elaborate orchestration.
Test the constraint in your system
The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check