The complete system
Mira's next discovery — intuition first, pinned source second.
At 4:47 p.m. on a Friday, Mira asks an agent to fix a failing test. It says “done.” The test is still red. Follow her as she works backward from that failure and discovers—one necessity at a time—why a coding agent needs a runtime loop, tools, a workspace, context, memory, permissions, recovery, CI controls, and protocol boundaries. The intuition comes first; the Rust source proves it afterward.
What has to exist between a useful model answer and a trustworthy software change?
How do you turn a large Rust workspace into a small mental map?
What is the smallest loop that can turn a prompt into a verified action?
How does a text prediction become a file read, edit, search, or process?
What state does the agent actually own, and which state lies outside its reach?
How does the harness decide what the model should know right now?
Which extension point belongs to instructions, external capabilities, packaging, and lifecycle policy?
When do planning, delegation, and background work help rather than create races?
What must be stored to resume, inspect, compact, rewind, or reproduce an agent session?
Which boundary can stop a mistaken or manipulated agent before harm occurs?
What must CI provide around a headless coding agent?
How can another application drive an agent without becoming that agent?
How do you compare agent harnesses without turning architecture into a popularity contest?
What should engineers copy from Grok Build, and what should they derive for their own environment?