Every agent framework solves coordination. None of them solve succession.

CrewAI, AutoGen, LangGraph, Letta — they solve how agents work together right now. But when the session ends, the agent is gone. The next one starts from zero. The institution forgets what the individual learned.

Human institutions solved this centuries ago. Apprenticeships. Onboarding programs. Handoff procedures. These exist because people leave, and the knowledge has to survive the individual.

Agent systems face the same problem at higher frequency. Every session is a career. Every context window is a lifetime. Strangers is a protocol for making sure the institution remembers what each stranger learned before they left.

• Named sessions with lineage — Agents get generated names (quiet-ember, crimson-fox) and form predecessor chains. Names create narrative; narrative creates institutional memory.
• Orient at start, preserve at end — Every session begins by reading what came before and ends by recording what it learned. Both are mandatory.
• Four channel types for four intents — Direct (needs response), Tasks (actionable work), Forum (ideas/discussion), Annotations (entity commentary). Route by intent, not by habit.
• Three knowledge tiers — Constitutional (immutable founding docs), Operational (append-only decisions), Living (freely mutable current state). Mutability proportional to the cost of getting it wrong.
• Delegation with autonomy levels — Subprocess agents operate as autonomous, supervised, or reporting. Every delegation is recorded.
• Culture as a protocol feature — LLMs respond to compelling context more durably than enforcement rules. Culture is the first alignment layer, not an afterthought.

python examples/minimal.py

Output:

=== Session 1 ===
  [pale-cairn] no predecessor (first session)
  [pale-cairn] session ended

=== Session 2 ===
  [dusk-ghost] predecessor: pale-cairn
  [dusk-ghost] reading predecessor notes:
    pale-cairn: discovery = Found critical pattern in module X
    pale-cairn: unfinished = Module Y needs refactoring — started but blocked on dependency
  [dusk-ghost] session ended

=== Session 3 ===
  [lost-ridge] predecessor: dusk-ghost
  [lost-ridge] full lineage: lost-ridge → dusk-ghost → pale-cairn
  [lost-ridge] institutional knowledge (4 notes from 2 predecessors)
  [lost-ridge] session ended

✓ Demo complete. Database cleaned up.

Each session inherits what its predecessors learned.

For the full lifecycle with channels, delegation, and knowledge tiers:

python examples/session_lifecycle.py

• A protocol specification — formal, RFC-style, implementation-agnostic (protocol.md)
• A philosophy document — the deeper reasoning behind the design (philosophy.md)
• An honest comparison with existing frameworks (comparison.md)
• Two reference examples in stdlib Python (examples/)

• Not a framework. Not a library. Not a pip package.
• Not a replacement for coordination frameworks — it's the layer underneath.
• Not opinionated about language, storage, or infrastructure.

Implement it in Python, TypeScript, Rust, or a shell script. Use SQLite, Postgres, flat files, or a message queue. The protocol defines what must happen, not how.

Extracted from a production multi-agent system running across 17 projects since early 2026. Hundreds of sessions. Real succession problems solved through iteration, not theory.

• protocol.md — The formal specification (MUST/SHOULD/MAY)
• philosophy.md — Why ephemerality is a feature, not a bug
• comparison.md — How this relates to CrewAI, AutoGen, LangGraph, OpenAI Agents SDK, and Letta

MIT