Blog
Benchmarks and engineering notes on agent tool surfaces. Everything here is backed by our public benchmark: 60 tasks, 1,800 judged trials, two agent models, six live servers.
How to combine multiple MCP servers into one endpoint
Client tool limits, duplicated config, and overlapping tools all push toward a single endpoint. The options, the tradeoffs, and what a compiler adds over a plain gateway.
2026-09-01
How to cut MCP token usage 5x without losing accuracy
Tool definitions are resent on every request and results flood the context. Five compiler passes that took a benchmarked agent from $0.260 to $0.052 per completed task.
2026-08-28
Too many MCP tools make agents worse. Here is the data.
Every connected MCP server adds tool definitions to every request. We measured what 137 tools across six servers does to agent accuracy and cost, and what fixes it.
2026-08-24