How Codex runs an agentSVGPNG
How Codex runs an agent HOW CODEX RUNS AN AGENT Context in. Decisions out. Tools act. Results return. CODEX AGENT Model + harness + tools HARNESS One controlled execution loop PERSISTENCE Read when needed PLUGIN Skills + connections LangGraph Orchestrate agents optional Agent A Agent B Agent C User Desktop / CLI / SDK Human or program entry Context Inputs for this model call 1 GPT / LLM Reason + choose the next action 2 Executor Validate + run tool requests 3 Policy Permissions + sandbox limits Compaction Instructions Goals, rules, AGENTS.md Skill Task methods + scripts MCP config Register external tools History Past messages + tool events Memory Recall across tasks State Params, job IDs, checkpoints TOOLS Callable interfaces MCP Client Connect using the protocol Local tools Run code + access files App / Connector Connect external accounts MCP Server Expose tools + resources Files / Shell / HPC Where commands actually run Services / APIs Data + external actions MCP Flow Load / Register Concept map Task Submit task infer Tool request Result gate retrieve load guide register Call / return read / write Final answer Access systems Summarize history MODEL: decides HARNESS: controls TOOLS: act