Technical explanations of how Orca schedules, isolates, observes, and connects production agents. Written from the implementation, with limits included.
Orca Harness is the execution kernel behind the Orcacode CLI and the Marlin runtime on Orca. What an agent harness actually does, why the loop is kernel code, and the measurements: 100-call fan-out under 0.3 ms, a 3.3 ms cold start, a 6.7 MB binary, and a live four-harness benchmark.
Read article
Orcacode is our terminal coding agent, built on the Orca Harness kernel. One curl command installs a single static binary for macOS, Linux, or Windows that finishes the same workload on roughly half the tokens of the leading harnesses, and opens itself to Agent Plugins, MCP, and an embeddable Rust SDK.
Read article
Every platform today hosts agents as deployed applications. Reasoning and execution have different infrastructure properties, and the next platform layer separates them into distinct planes.
Read article
Orca’s MCP bridge keeps full tool schemas server-side and shows the model a small card on demand. In a stress test, one tool’s schema dropped from an estimated 54K tokens to 1.3K, a 97.7% cut, with exact execution intact.
Read article
MCP security isn't an allowlist. It's five checks between the model and the tool: an opaque handle, a pinned schema, a risk gate, a signed bridge and an egress check. Here's what each one stops, with the numbers from our source.
Read article
AI agent cost control needs a hard boundary before each call, not a dashboard after the bill. Use per-run budgets, reservations, and clear stop reasons.
Read article
An MCP timeout tells you the caller stopped waiting, not what the tool did. Give every call a deadline, classify the error before you retry, and never repeat a write you can't prove never happened.
Read article
Claude Code headless mode in production: permission allowlists, JSON output with cost per run, cron and launchd scheduling, and the failure modes we hit.
Read article