DNA / Context Architecture
Context Architecture
Every agent knows exactly where it is.
Context engineering is the discipline of designing information structures that allow AI agents to operate effectively with minimal ambiguity. SESSION-ANCHOR.md, brain.md, and AGENTS.md form a three-layer context system — project state, project intent, and project rules.
Key Principles
- 01SESSION-ANCHOR.md: compressed state map — any agent can onboard cold in under 60 seconds
- 02brain.md: project context, architectural decisions, and intent per repository
- 03AGENTS.md: guardrails, SOPs, and workflow rules that constrain agent behavior
- 04Context is an asset that compounds — better context = fewer tokens wasted on discovery
Real Examples from Production
agent-os SESSION-ANCHOR
Contains current phase, commands to run, graph pipeline state, and next step. One file replaces a 30-minute onboarding.
agent-os/SESSION-ANCHOR.md →
jadzia-core brain.md
Records architectural decisions (why FastAPI, why JWT, why Paramiko) so future agents don't re-debate them.
jadzia-core/brain.md →
AGENTS.md in every repo
8 repositories, 8 AGENTS.md files. Each one is a behavioral contract for every AI agent that enters that codebase.
agent-os/AGENTS.md →
Traditional vs AI Systems Architect
Traditional Developer
Developer reads README, explores codebase, asks questions for 2 hours before making a change.AI Systems Architect
Agent reads SESSION-ANCHOR.md (90 seconds), knows exact state, runs prescribed command, delivers result.