The fastest way to break a careful architecture decision is to hand a fresh agent a file it has never seen. It can't read your mind, and it can't read a chat log that no longer exists — so it optimizes for the local change and quietly undoes a constraint you fought hard to establish.
Every new session starts from zero
Consider a session manager that deliberately rotates refresh tokens and invalidates the old one on use. It looks like extra complexity. A new agent, asked to "simplify the refresh flow," may helpfully remove exactly the line that closes a security hole — because nothing told it the line was load-bearing.
Ask before you edit
Kodela gives agents a retrieval step: before modifying a file, an agent calls get_context and receives the decisions and reasoning attached to it. The relevant decision — "refresh tokens must rotate; old token invalidated on use" — becomes part of the agent's working context, not a tribal secret.
- The agent sees human-authored decisions before it touches the code they govern.
- Load-bearing constraints are explicit, so simplification doesn't silently weaken them.
- Onboarding a new tool — Claude Code, Cursor, anything MCP-capable — inherits the same memory.
The decision didn't live in someone's head anymore. It travelled with the file.
A codebase that explains itself to machines
This is the quiet promise of context infrastructure: the same memory that helps a human reviewer also steers the next agent. As more of your team's work is delegated to AI, the decisions that keep the system coherent stop depending on whoever happened to be in the room.