Skip to content

Agent capabilities

Which UI controls appear for a given agent is derived from what that agent's adapter actually supports, not a single shared feature set.

Claude Code, Codex, and Antigravity don't expose the same capabilities, so AgentDock doesn't pretend they do. Each adapter declares what it supports, and the UI shows or hides controls accordingly rather than showing a control that would silently no-op.

What varies by agent

CapabilityClaude CodeCodexAntigravity
Live model switching mid-sessionYesYesNo — model changes restart the underlying process
Live permission-mode switching mid-sessionYesNo — fixed at thread creationNo — fixed at launch
Structured, typed event transportYes (SDK)Yes (SDK)No — PTY output classification
Image/attachment supportNo native supportYesYes
Reasoning-effort controlWhere the SDK exposes itWhere the SDK exposes itWhere the CLI exposes it

See Claude Code, OpenAI Codex, and Google Antigravity for how each adapter is actually implemented.

Why this matters for Compare Mode and AIgency

Compare Mode and AIgency both run real sessions across multiple agents at once, so the same capability differences apply there too — for example, Compare Mode's forced read-only modes are expressed differently per agent (plan for Claude Code, read-only for Codex, plan for Antigravity) precisely because each agent's permission model is different. See Workspace behaviour and Agent assignment.