# Making Drugverse Agent-Accessible A long-horizon agent works for hours or days. It needs stable IDs, pinned versions, cheap outputs and async jobs. A chat connector alone does not give this. ## What long-horizon agents need ### 1 · One core, three surfaces - MCP server for ChatGPT, Claude and agent frameworks. See [[Building ChatGPT Apps and Claude Connectors]]. - CLI for coding agents. Agents use CLIs well. See [[CLI vs MCP]]. - REST API for pipelines. - All three call the same service layer over Postgres and the knowledge graph. ### 2 · Identity resolution first - Agents guess gene symbols and drug names. They get them wrong. - `resolve_identifier` is the first call in every workflow. It returns one canonical ID and flags ambiguity (old symbol, gene vs protein, renamed drug). - This is Scope's disambiguation, exposed as a tool. It is the strongest agent feature. ### 3 · Pinned versions - Every call accepts `as_of` (a Drugverse release ID). - The agent pins one release for the whole run. Answers stay consistent across hours. - Every response returns the release ID. ### 4 · Provenance in every answer - Each fact carries: source, source ID, source version, evidence link. - The agent can cite. A human can audit. A reviewer can reproduce. ### 5 · Token-cheap outputs - Return summaries and **handles**, not full dumps. Example: "412 targets found, handle `rs_81`. Top 10 below." - Let the agent filter, sort and page on the server. - Hard row limits and timeouts on graph and SQL tools. ### 6 · Async jobs - Heavy analysis (repurposing scan, graph refresh) runs as a job. - Return a job ID. The agent polls or gets a callback. Jobs are idempotent. - MCP Tasks support this pattern (experimental since the 2025-11-25 spec).^[Source: [MCP 2025-11-25 changelog](https://modelcontextprotocol.io/specification/2025-11-25/changelog)] ### 7 · Workspace memory - Saved result sets and notes per project. - An agent resumes after a restart without redoing queries. ### 8 · Agent docs - `llms.txt` and tool descriptions with worked examples. - Skills that encode workflows: "repurposing analysis", "target dossier". See [[Agent Skills as Codified Domain Expertise]]. ### 9 · Security - Read-only scopes by default. Per-key rate limits. Audit log of every call. - **DB Parser Pro**: ship the same MCP server inside the client network. Private data never leaves. ## Build order 1. Service layer + `resolve_identifier` + `get_entity`, with `as_of` and provenance. 2. MCP server (read-only tools). Test in Claude and ChatGPT. 3. CLI on the same core. 4. Handles, paging and async jobs. 5. Workspace memory and skills. ## Test with the RL tasks - Reuse the tasks from [[Drugverse RL Environments and Data Play]] as an agent regression suite. - The same tasks then serve as product QA and as a sellable environment. ---- tags: #boat #ai links: [[Drugverse]] [[Context Layers MOC]] [[Agent Harnesses and Agentic Search]] [[AI Agents Stack]] [[knowledge graphs]]