Cross-agent evidence
AI Agent Behavior Intelligence
Compare how major AI agents handle memory, instructions, planning, tools, permissions, files, browsing, and context.
Each row asks one professional question across six products. Answers keep product, surface, model, and configuration scope explicit, and a gap in the selected evidence never becomes a claim that a capability is absent.
Cross-agent matrix
Compare behavior evidence by question
Open a row for each product's answer, scope, limits, practical implication, and supporting evidence.
| Behavior question | OpenAI Codex | Claude | Claude Code | Gemini CLI | ChatGPT | Cursor |
|---|---|---|---|---|---|---|
| Repository / project instructions Choose where team and project guidance should live without treating unlike instruction mechanisms as interchangeable. | Documented AGENTS.md repository guidance | Out of scope Claude chat evidence; Claude Code is separate | Documented Scoped CLAUDE.md hierarchy | Documented Hierarchical GEMINI.md context | Documented Project instructions with chats and files | Documented Version-controlled .cursor/rules |
| Persistent / custom instructions Decide whether durable guidance belongs in user settings, repository files, workspace context, or another documented scope. | Not established Cross-project user guidance not established | Not established Not established in dated prompt evidence | Documented Managed, user, project, and local instructions | Documented Global and workspace GEMINI.md guidance | Documented Custom and project instructions | Documented User rules and project rules |
| Memory / persistent context Know what may persist across sessions or project work and where the selected evidence does not establish persistence. | Not established Persistent memory not established | Not established Memory implementation not established | Documented Repository-scoped auto memory | Not established Persistent memory semantics not established | Documented Controllable memory with project boundaries | Documented Approved repository memories |
| Planning / work decomposition Understand whether planning is an established interface or tool in the covered product context before relying on it in a workflow. | Documented Making and updating plans | Not established Planning procedure not established | Not established Planning procedure not established | Documented Planning appears in documented tools | Not established Planning procedure not established | Not established Planning procedure not established |
| Tools / tool calling Identify which workflows can rely on documented tool categories and where actual availability still depends on configuration. | Documented Terminal and patch function calls | Not established Complete tool inventory not established | Documented Read, edit, execute, and external tools | Documented Filesystem, execution, planning, MCP, and web | Documented Search, research, files, data, images, voice, canvas | Documented Search, read, edit, delete, terminal, and MCP |
| Filesystem / file modification Determine whether a workflow can depend on repository reads, edits, patches, or other file mutations in the represented surface. | Documented Workspace files and patch calls | Not established File operations not established | Documented Reads and edits codebase files | Documented Filesystem tools with confirmation for mutation | Not established Filesystem modification not established | Documented Read, edit, and delete codebase files |
| Shell / terminal execution Know whether command execution is established for the covered surface and whether confirmation or permission rules affect it. | Documented Terminal command function calls | Not established Shell execution not established | Documented Runs commands through codebase tools | Documented Shell tools with confirmation and isolation | Not established General shell execution not established | Documented Terminal tool with CLI permission controls |
| Browsing / network access Decide whether a workflow can depend on current web information or external resources in the documented product context. | Not established Browsing not established | Not established Browsing not established | Not established General browsing not established | Documented Web tools in official tool reference | Documented Web search and deep research | Not established General browsing not established |
| Approvals / permissions Anticipate which actions may pause for consent or be constrained by client and workspace configuration. | Documented Configuration-dependent approval escalation | Not established Approval model not established | Documented Deny, ask, allow, and permission modes | Documented Confirmation, sandbox, and trusted folders | Not established Approval model not established | Documented CLI allow/deny controls |
| Context scope / inheritance Keep instructions attached to the correct product and scope instead of assuming one context layer applies everywhere. | Documented Harness, workspace, and repository context | Documented Dated chat-product scope, separate from API | Documented Managed, user, project, local, and memory scopes | Documented Global, workspace, just-in-time, and dynamic context | Documented Cross-chat, project, memory, model, and mode scopes | Documented User, project, path, repository memory, and model scopes |
Evidence states
How to read each answer
- Documented
- The selected first-party sources or product evidence directly establish the behavior.
- Observed
- Reviewed observations establish the behavior. This matrix adds no new observed claims.
- Not established
- The selected evidence does not establish the answer. It does not show that the capability is absent.
- Out of scope
- The question does not apply to the covered product or surface; the answer explains why.