There's a pattern in how agent platforms fail that has nothing to do with which model they run. You connect everything — the code host, the chat tool, the tracker, the wiki — and the agent gets worse. Slower to pick the right action, more likely to wander mid-task, worse at the thing it was good at last week when it had less.
That isn't the model getting dumber. It's the context window filling up with things that aren't the task.
The part you can measure
Before an agent reads the first word of your request, its context window already contains the definition of every tool it might use — names, descriptions, parameter schemas. Connect a handful of servers and that overhead stops being a rounding error. Anthropic's own engineering write-up on code execution with MCP walks through this failure mode: tool definitions loaded up front, intermediate results dragged through the window, and most of the budget spent before the work starts.
One enterprise case study reported three connected servers — roughly forty tools — consuming about 143,000 of a 200,000-token window before the first user token, and tool-selection accuracy collapsing from 43% to under 14% as the toolset bloated. Researchers call the broader effect context rot: the measurable accuracy drop as input grows. Whatever the exact numbers on your stack, the direction is not in dispute — every token of tool manual is a token the model isn't spending on your task.
And long-horizon work compounds it. A two-hour run makes hundreds of decisions; a small per-decision degradation is the difference between an agent that finishes and one that wanders.
The design decision
The industry's default answer is additive: more tools, more connectors, a bigger catalog on the pricing page. The number goes up and to the right, and every agent pays for the whole catalog on every single call.
We made the opposite bet. On Auteryn, the platform has 90+ built-in tools — but no agent carries all of them. An agent carries exactly what you gave it:
- Capabilities are per-agent. The browser, media generation, computer use — each is a switch, not a default.
- Connections are chosen, and bounded: an agent writes through exactly one connection per integration. That's enforced in three layers, not by convention.
- Skills load progressively. An agent sees a one-line catalog; the full instructions load only when a skill is actually used — not before.
A support agent might carry nine tools. A coding agent might carry thirty. Neither pays for the other's toolbox, and neither pays for the seventy tools it will never call.
The bonus nobody markets
Focus turns out to be a security property. An agent that was never given a tool cannot call it, leak through it, or be talked into using it by a hostile web page. Least privilege has been the boring backbone of security for fifty years; applying it to the agent itself is the same idea with new stakes. Our security page states it plainly: the smallest possible tool surface, per agent.
Ten focused agents
The objection writes itself: "but I want one agent that does everything." So did we, for about a month. What we found — building this platform with its own agents — is that everything is exactly what a team is for. Ten focused agents on one serious harness beat one agent that does everything badly, for the same reason ten specialists beat one exhausted generalist: each one shows up with a clear head.
The harness underneath is identical — the same machine, the same checkpoints, the same memory, the same approval gates. Focus costs nothing. It's just the refusal to make every agent carry everyone's luggage.
If you want to see what a focused agent feels like, the free plan is a real agent on a real sandbox — 1,000 credits a month, no sales call.