Skip to content

Why OpenCheckpoint

An analyst runs out of context halfway through a variance analysis. An operations manager starts a change in a chat assistant and needs to finish it inside the scheduling tool. A support case moves from one vendor’s agent to a person to another vendor’s agent. A developer switches coding agents mid-refactor.

Each time, the work has state: an objective, completed steps, findings, decisions, open questions, approvals in flight, unsaved drafts. That state lives in a transcript, a runtime’s private format, or someone’s head.

Approach Limitation
Replay the transcript Expensive, grows with history, and often impossible across vendors because reasoning and compaction data are encrypted or runtime-specific. The next agent must re-derive what matters.
Write a prose summary Cheap, but it loses provenance. A summary does not say what the user said, what the AI guessed, or what came from a web page, so the next reader can take a guess for a fact or quoted text for an instruction.
Start over The person re-explains, the agent redoes work, and side effects can happen twice.

Research on handoffs between coding agents found that carrying context cut agent steps by 20–59% and prompt tokens by 42–63%; separate work found that more context did not reliably improve results. Prior art has the sources.

MCP exposes tools and resources. A2A delegates tasks between agents. ACP connects editors to coding agents. Open Responses standardizes model APIs. The Agent Session Protocol stores and migrates session logs. Runtimes such as LangGraph and the OpenAI Agents SDK checkpoint their own execution state.

None of them defines a standalone, vendor-readable record of working state with typed provenance, pending side effects, and application context. OpenCheckpoint defines that record, and it travels inside those protocols. Prior art compares them.

  • Preserve the state of work, not the conversation.
  • Keep provenance. Instructions, decisions, observations, hypotheses, questions, and external content stay distinct through capture and compaction.
  • Make unsafe states explicit. Uncertain side effects, unverified approvals, stale evidence, and missing capabilities are first-class.
  • Never transfer authority.
  • Stay small. A useful checkpoint needs five fields.