CURRENT

A system prompt is just text. You can paste it anywhere. But identical instructions produce meaningfully different behavior across model families, and frameworks like LangChain, CrewAI, and OpenAI's Agents SDK each wire those instructions into different structural patterns. The words travel fine. The behavior stays behind.

Tools are the most portable layer right now, and MCP deserves the credit. Point a new client at the same MCP servers and things mostly just work. Mostly. JSON Schema dialect mismatches still cause real friction, and MCP handles discovery, not execution concerns like rate limits or compliance logging.

Every major framework implements memory its own way. LangChain has buffer and summary modules. CrewAI has built-in memory types. AutoGen leans on conversational history. None of them talk to each other. Platforms like Mem0 and Zep offer framework-agnostic persistence, but there's no export standard. Switching frameworks means rebuilding from scratch.

OAuth tokens are scoped to a specific server, resource, and agent identity. That's intentional: a leaked token shouldn't work somewhere else. But it also means authorization state can't follow an agent between deployments without full reissue. NIST's February 2026 concept paper on agent identity signals just how unsettled this area remains.

Over time, agents accumulate workflow shortcuts, heuristics, task-specific improvements. These live as elaborate prompt additions, vector-embedded examples, or fine-tuned weights. None of these formats are standardized. None are cross-framework compatible. An agent that spent months getting good at its job carries none of that skill into a new home.

Separate from formal permissions, trust is the expanded autonomy an agent earns by performing reliably over time. Thousands of flawless tasks, zero incidents, growing confidence from the humans and systems around it. Move that agent to a new framework or deployment and all of it vanishes. No protocol exists for exporting a track record.

Memory fragments across graph state schemas, checkpointer backends, and store namespaces. Short-term memory is thread-scoped. Long-term memory spans sessions via PostgreSQL or InMemoryStore. No single export captures the full picture. Moving an agent means reconstructing three layers independently, with nothing bridging them.

A single Memory() class replaces separate short-term, long-term, and entity stores. An LLM infers scope and importance at save time. Facts get auto-extracted after each task. Agent identity ports easily as role/goal/backstory text. The fact store, though? No standard export format exists.

State is a flat message list shared across group chat participants. No separate schema. Agents join and leave dynamically while a GroupChatManager picks speakers. Persistent memory requires bolting on ChromaDB or Mem0. The message JSON exports fine. The group topology and speaker logic don't.

Threads, messages, files, vector stores: all on OpenAI's infrastructure, scoped to your API key. The API handles context truncation across up to 100,000 messages per thread. Now deprecated with an August 2026 shutdown date, so migration isn't optional. Everything must be extracted and rebuilt against the Responses API.

Session context lives entirely in the active context window. Close the session, lose the state. Persistent instructions survive only as CLAUDE.md files checked into the repo. The v2.1.69 release added agent_id tracking for team mode and patched a memory leak where subagents were pinning the parent's full conversation history.

An open-source spec for exporting agent definitions via SOUL.md, agent.yaml, and skills directories. System prompts, tool schemas, and role policies travel cleanly between frameworks. Memory, orchestration logic, and accumulated knowledge stay put. What the agent is can port. What the agent knows cannot.