Keeping an agent fleet useful at scale
Supervising a fleet of coding-agent sessions requires more than a dashboard. Teams need current context, evidence that routing choices hold up, and a way to use lower-cost capacity without giving up control. claude-deck brings those concerns into a native desktop cockpit with live observability and a cost-aware model-routing layer.
Architecture highlights
- A router that picks the cheapest capable model for each task, with confidence and spend limits blocking bad downgrades.
- An evaluation harness that grades agent output: LLM-as-judge, head-to-head model comparisons, real-repository task banks.
- MCP fleet memory and a semantic cache keyed by repo and commit, so sessions reuse each other’s work instead of re-deriving it.
- A multi-provider coding agent built from scratch with tool execution, retries, and speculative prefetch.
Verified scale
- 157K Rust product LOC across an 8-crate workspace.
- 4,571 test functions and 2,484 commits.
- On a 30-task benchmark, the router kept 21 tasks on a free local model instead of paid APIs.
Availability
The code is private; a walkthrough is available on request.