Google AI product-system dossier 064

Google AI Developer and Agent Platform: Architecture, Tools and Production System

How AI Studio, Antigravity, Gemini CLI, Gemini Code Assist, Jules, the Gemini API, ADK, Agents CLI and Google Cloud's managed Agent Platform fit together.

Published 24 August 2026 · 20:45 ICT · information current to 24 August 2026 · 20:45 ICT

Part I — System analysis. Sections 01–14 explain the product boundaries, execution loop and route from prompt to production agent. Part II records the current command surfaces separately so that interface controls are not confused with architecture.

The platform in one minute

Google's developer-agent system is a portfolio, not one application. Gemini models provide reasoning; the Gemini API exposes model and tool interfaces; AI Studio, Antigravity, Gemini CLI, Code Assist and Jules provide different build or coding surfaces; ADK supplies an open agent framework; and Google Cloud's Agent Platform supplies managed runtime, memory, evaluation, observability and governance.^[1][2][3][4]

Google AI Studio is therefore not a complete equivalent to Claude Code or Codex. It is the simplest browser route from prompt experimentation to a Gemini application. Antigravity is the closest Google surface for supervised local agentic work; Gemini CLI is the open-source terminal agent; Code Assist is the IDE and enterprise coding layer; Jules is the asynchronous GitHub worker. The products overlap, but their execution, context, permissions and deployment boundaries differ.

Model and APIGeminiReasoning, generation and tool requests

Interactive surfaces5AI Studio, Antigravity, CLI, Code Assist, Jules

Agent frameworkADK 2.0Open, multi-language and model-flexible

Production control planeAgent PlatformRuntime, state, evaluation and governance

Canonical formula. Production agent = model + instructions + context + tools + execution host + permissions + session/state/memory + evaluation + observability. A model can propose an action; only the host application or managed runtime can authorise and execute it.

Essential questions

Question	Concise answer
Is Google AI Studio the equivalent of Claude Code or Codex?	No. AI Studio is primarily a browser environment for Gemini prompts and app construction. Antigravity, Gemini CLI, Code Assist and Jules cover the agentic coding roles that Claude Code or Codex combine more closely.
What is Google's closest direct coding-agent equivalent?	There is no single equivalent. Antigravity is closest for local multi-agent work; Gemini CLI for an open terminal workflow; Code Assist for IDE work; and Jules for asynchronous repository tasks.
What is the difference between Gemini and the Gemini API?	Gemini is the model family. The Gemini API is a developer interface that sends content to a selected model and exposes generation, structured output, tool calling and other capabilities.
How do Gemini CLI and Antigravity CLI differ?	Gemini CLI is an Apache-licensed open-source terminal agent with its own extensions and configuration. Antigravity CLI shares Google's newer Antigravity agent harness and settings with the desktop command centre.
What is Gemini Code Assist?	Google's coding product for supported IDEs and enterprise development workflows. Its agent mode can plan, inspect files, call tools and request permission for mutating actions; the mode remains labelled Preview in the current documentation.
What is Jules?	An asynchronous GitHub-connected coding agent. Each task runs in a Google Cloud virtual machine, clones the selected repository, proposes a plan and returns logs, diffs and changes for review.
What is ADK?	The open Agent Development Kit: a framework for composing agents and deterministic workflows with tools, sessions, state, memory, callbacks, plugins, evaluation and deployment adapters.
What is Agents CLI?	A lifecycle tool for ADK projects. It scaffolds, tests, evaluates, packages, deploys and observes agents, and can install agent-building skills into supported coding agents. It is not the model or production runtime.
What is Google Cloud's Agent Platform?	The managed production layer formerly described through Agent Builder and Agent Engine names. It provides runtime, sessions, memory, evaluation, code execution, computer use, tracing, logging and enterprise controls.
What is the difference between MCP and A2A?	MCP standardises access to tools and context. Agent2Agent enables independently built agents to discover, communicate and collaborate. They solve adjacent, not identical, integration problems.
Who actually executes a tool call?	The model returns a tool name and arguments. The calling application, agent harness or managed runtime validates policy, executes the tool and returns the result to the model.
Can ADK use non-Gemini models?	Yes. ADK is optimised for Google's stack but its official documentation supports multiple providers and deployment environments. Model portability does not make every tool or managed service portable.
Where does work execute?	AI Studio uses hosted build services; Antigravity and Gemini CLI can act on a local workspace; Code Assist acts through an IDE; Jules uses a cloud VM; ADK agents run wherever deployed; Agent Platform is managed Google Cloud.
How are context and memory different?	Context is the evidence supplied for the current model call. A session groups one interaction thread; state holds working values; memory makes selected information searchable across sessions.
What is the central security boundary?	Separate model reasoning from execution authority. Instructions do not grant credentials, a proposed call is not approval, and a successful tool response is not proof that the overall task is correct.
What should an enterprise standardise first?	Identity, tool allow-lists, credential scope, data boundaries, evaluation sets, audit logs, deployment promotion and human approval for consequential actions—before maximising agent autonomy.

The system and its layers

Canonical architecture. The layers can be combined in several ways. AI Studio can build directly on the Gemini API; an ADK agent can run locally or on a managed runtime; a coding surface can itself call MCP tools or generate an ADK project.

Layer	Responsibility	Google components	Do not confuse it with
Surface	Accept instructions, show plans and review outputs.	AI Studio, Antigravity, Gemini CLI, Code Assist, Jules.	The model or production runtime.
Model	Reason, generate and propose tool calls.	Gemini model catalogue.	Permission to execute.
API	Transport content, configuration and tool schemas.	Gemini API and Vertex AI model endpoints.	An autonomous agent.
Framework	Define agent composition, workflows and lifecycle hooks.	ADK and Agents CLI.	A hosted application surface.
Integration	Connect tools, context and other agents.	Function calling, MCP and A2A.	Identity or governance.
Execution	Perform side effects under host policy.	Local process, IDE, Jules VM, Code Execution, Computer Use.	The model's textual intention.
Production	Host, persist, evaluate, observe and secure agents.	Agent Platform, Cloud Run and GKE.	The development framework alone.

From model API to production agent stack

Date	Milestone	What changed	System meaning
6 Dec 2023	Gemini and AI Studio	Google introduced the Gemini model family and a browser route for developers to test prompts and obtain API code.	Established the model/API and lightweight developer-surface layers.
Apr 2024	Gemini Code Assist	Google's IDE coding assistance adopted the Gemini name and expanded enterprise development support.	Placed models inside the software-development lifecycle.
9 Apr 2025	ADK and A2A	Google released an open agent framework and proposed an interoperability protocol for agents.	Separated agent construction from any single user interface or model.
20 May 2025	Jules public beta	Asynchronous repository tasks ran in isolated Google Cloud virtual machines.	Added delegated cloud coding with reviewable plans and diffs.
25 Jun 2025	Gemini CLI preview	An open-source terminal agent exposed Gemini, built-in tools and extensibility at the command line.	Created a local, scriptable coding-agent route.
6 Aug 2025	Jules general availability	Jules moved beyond beta and expanded task and repository workflows.	Made asynchronous coding a durable product surface.
8 Oct 2025	Gemini CLI extensions	Packaged prompts, tools and MCP integrations became distributable extensions.	Broadened the terminal agent beyond its built-in tools.
19 May 2026	Antigravity 2.0 and managed agents	Google advanced its multi-agent development environment, AI Studio build harness and managed-agent interfaces.	Joined interactive development with a clearer route to production agents.
2026	ADK 2.0, Agents CLI and Agent Platform	Agent graphs, lifecycle tooling, evaluation, deployment and managed runtime services matured as a coherent stack.	Shifted emphasis from demonstrations to governed operational systems.
24 Aug 2026	Research cut	The portfolio still contains overlapping generations, names and maturity labels.	Documentation for the selected surface remains the live authority.

Chronology boundary. A launch date does not mean that every feature is generally available. Code Assist agent mode is still documented as Preview; managed services can have region, model, framework or control differences.

Choose the surface by operating model

Surface	Primary role	Execution and context	Best fit	Maturity boundary
Google AI Studio	Prompt experimentation and agent-built web or Android applications.	Hosted build environment; React + Node or Kotlin + Compose; GitHub and Cloud Run routes.	Fast Gemini prototypes and compact full-stack applications.	Not a repository-wide terminal agent or production control plane.
Antigravity desktop	Local agent command centre for coding and knowledge work.	Workspace files, shell, browser, tools, subagents, skills and scheduled work under permissions.	Supervised multi-agent work across a local project.	Local configuration and permissions remain decisive.
Antigravity CLI	Terminal access to the Antigravity harness.	Multi-file edits, tool calls and shared Antigravity configuration.	Developers who prefer a terminal agent with the newer harness.	Not the same codebase or extension system as Gemini CLI.
Gemini CLI	Open-source terminal agent and automation interface.	Local tools, GEMINI.md context, MCP, extensions and non-interactive commands.	Transparent, scriptable and community-extensible terminal workflows.	Capabilities depend on version, authentication and extensions.
Gemini Code Assist	IDE assistance and enterprise repository workflows.	Editor context, files, tools, MCP and organisation configuration.	Interactive coding inside VS Code or JetBrains-family IDEs.	Agent mode is Preview; editions and admin controls differ.
Jules	Asynchronous GitHub issue and code tasks.	One hosted VM per task, cloned repository, plan, logs and diffs.	Delegated background maintenance and bounded feature work.	Repository access and review remain explicit user responsibilities.
ADK + Agents CLI	Build, test and package custom agents.	Application code plus model, tools, sessions, workflows, evaluation and deployment adapters.	Teams creating an agent product rather than using a coding assistant.	A framework does not supply production operations automatically.
Agent Platform	Operate agents in production.	Managed runtime, sessions, memory, evaluation, tracing, security and gateways.	Governed cloud services and enterprise agent fleets.	Service availability and compliance scope must be checked per workload.

How an agent action actually happens

Execution sequence. Google's function-calling documentation states the key boundary directly: the model predicts a function call; the application executes it. Agent products automate more of the loop but do not erase the separation.

Stage	Input	Decision owner	Output	Failure mode
Context assembly	User request, history, files, instructions and tool schemas.	Surface or agent runtime.	Bounded model input.	Missing, stale or malicious context.
Reasoning	Current context and prior observations.	Selected Gemini model.	Text, structured output or proposed call.	Incorrect plan or fabricated premise.
Authorisation	Tool, arguments, policy and user settings.	Host/runtime and, where required, human.	Allow, ask or deny.	Overbroad allow rule or approval fatigue.
Execution	Approved action and credentials.	Tool host or managed service.	Side effect and machine result.	Unsafe change, excessive privilege or partial operation.
Observation	Tool result, logs and errors.	Agent runtime.	New evidence for the model.	Truncated, ambiguous or adversarial output.
Verification	Claimed result and independent checks.	Agent plus reviewer under task policy.	Accepted result, revision or rollback.	Confusing successful execution with correct completion.

Instructions, context, sessions and memory

Object	Purpose	Scope	Persistence boundary
Prompt	States the immediate task and constraints.	Current interaction.	Retained only as the surface or session preserves it.
GEMINI.md	Supplies hierarchical project instructions to Gemini CLI.	User, workspace and nested directory scopes.	File-backed and versionable when stored with the project.
AGENTS.md	Supplies repository guidance where supported by the active agent surface.	Repository or directory tree.	File-backed; support and precedence are surface-specific.
Session	Groups one conversation and its events.	One user-agent thread.	Local in-memory services can lose it on restart; managed services persist by policy.
State	Holds working values associated with a session.	Application-defined keys and workflow steps.	Depends on the selected session service.
Memory	Makes selected information retrievable across sessions.	User, application or configured corpus.	Managed separately from the model context window.
Artifact	Stores files or binary outputs created by an agent.	Session, user or application design.	Requires an artifact service or external store.
Retrieved context	Adds repository, enterprise or web evidence at run time.	Query and connector permissions.	Source system remains authoritative.

Authority boundary. A project instruction file can shape behaviour but cannot create operating-system rights, Cloud IAM permissions or credentials. Instruction precedence, tool policy and execution authority are separate control planes.

Tools and execution boundaries

Tool class	Examples	Read or write	Primary control	Evidence required
Workspace	List, search, read and edit files.	Both	Workspace scope, path policy and review.	Diff plus targeted tests.
Shell	Build, test, format and repository commands.	Both	Command permissions and sandbox.	Exit code, logs and changed-file audit.
Web	Search, fetch and browser interaction.	Usually read; browser actions may write externally.	Domain, network and action permissions.	Direct source and action confirmation.
Function calling	Application-defined APIs with schemas.	Either	Host validation, credentials and business rules.	Structured response and domain check.
Code Execution	Run generated code in a managed environment.	Contained side effects	Runtime isolation, resource and package policy.	Inputs, outputs and reproducibility.
Computer Use	Observe and operate graphical interfaces.	Both	Action policy, environment isolation and human confirmation.	Screen state and external-system verification.
MCP tool	Connected database, SaaS or internal service.	Either	MCP server trust, exposed schema and downstream identity.	Server result plus source-system audit.

Skills, plugins, MCP and A2A

Mechanism	What it packages	When to use it	Trust boundary
Skill	Reusable instructions, workflows and supporting resources.	Repeatable domain or repository tasks.	Instructions can influence tool use; inspect before enabling.
Antigravity plugin	Skills, agents, rules, MCP configuration and hooks.	Distributable capability bundles for the Antigravity environment.	Bundle provenance and every included executable or connection.
Gemini CLI extension	Prompts, commands, MCP servers and configuration.	Portable terminal workflows and integrations.	Install source, update path and exposed credentials.
MCP	Standard tool and context interfaces.	Give a model-controlled client access to an external capability.	The server and downstream service retain independent permissions.
A2A	Agent identity, capability discovery, messages and task exchange.	Collaboration between independently built or hosted agents.	Remote agent identity, data disclosure and delegated action.
ADK plugin/callback	Lifecycle behaviour around agents, models, tools and events.	Policy, telemetry, transformation and reusable runtime logic.	Runs inside the application lifecycle and can alter behaviour.

MCP is not A2A. An agent can use MCP to operate a database tool and A2A to delegate a task to another agent. Neither protocol automatically supplies authorisation, audit, evaluation or safe data classification.

Agents, workflows and orchestration

Pattern	Control logic	Strength	Weakness	Google route
Single agent	One model loop selects among tools.	Simple context and ownership.	Long tasks can become opaque or fragile.	Gemini API, ADK LLM agent or coding surface.
Sequential workflow	Predetermined stages pass outputs forward.	Auditable and reproducible.	Less adaptive when inputs vary.	ADK graph/workflow agents.
Router + specialists	A coordinator assigns tasks to specialised agents.	Clear expertise and smaller contexts.	Routing and synthesis can fail.	ADK multi-agent or Antigravity subagents.
Parallel workers	Independent tasks execute concurrently.	Lower elapsed time for separable work.	Conflicts, duplicate work and higher spend.	Antigravity tasks, Jules tasks or ADK branches.
Human checkpoint	A plan, tool or promotion waits for approval.	Controls consequential transitions.	Latency and approval fatigue.	Code Assist permissions, Jules plan review, deployment gates.
Agent-to-agent	A2A messages cross agent boundaries.	Framework and vendor interoperability.	Distributed identity, trust and observability complexity.	ADK plus A2A-compatible agents.

From prototype to production

Lifecycle stage	Primary Google route	Production question	Required evidence
Explore	AI Studio, Gemini API or local CLI.	Can the model and tools solve representative tasks?	Recorded prompts, tool traces and failure examples.
Construct	ADK, Agents CLI or application code.	Are agent roles, workflows and state explicit?	Versioned code, schemas and deterministic tests.
Evaluate	ADK evaluation, Agents CLI and Agent Platform Evaluation.	Does quality hold across a defined dataset and grader?	Baselines, regression thresholds and error taxonomy.
Deploy	Agent Platform runtime, Cloud Run or GKE.	What is the identity, network and persistence boundary?	Infrastructure configuration and promotion record.
Persist	Sessions, State, Memory Bank and artifact stores.	What survives a request, session or restart?	Retention, deletion, tenancy and migration tests.
Observe	Tracing, logging, monitoring and custom telemetry.	Can operators reconstruct model and tool decisions?	Correlated traces, costs, latency and tool outcomes.
Govern	IAM, VPC controls, CMEK, data residency and Agent Gateway.	Who may call which model, agent, tool and data source?	Policy tests, audit logs and incident procedures.

Deployment boundary. ADK can run locally, on Cloud Run, on GKE or on a managed agent runtime. Portability at the framework layer does not imply identical sessions, memory, evaluation, security or operational behaviour across targets.

Security, permissions and governance

Control	What it governs	Practical rule	Residual risk
Instruction hierarchy	Which behavioural guidance takes precedence.	Keep trusted project rules short, scoped and versioned.	Prompt injection can enter through files, web pages or tool output.
Antigravity permissions	Deny, ask or allow command patterns; deny takes precedence.	Default mutating or high-impact commands to ask or deny.	Overbroad patterns and user habituation.
Sandbox	Technical containment of execution.	Treat it separately from approval policy.	A permitted action can still be logically harmful.
Identity and credentials	Cloud, repository and connected-service authority.	Use workload identity and least-privilege, short-lived credentials.	Downstream systems may expose broader rights than the agent needs.
Data controls	Residency, encryption, retention and service perimeters.	Verify the exact service, region, model and feature combination.	Preview features and third-party tools may sit outside the intended boundary.
Evaluation	Quality, policy and regression behaviour.	Test both expected tasks and adversarial inputs before promotion.	Static datasets miss new environments and attack methods.
Human approval	Consequential tool calls and deployment transitions.	Require explicit confirmation for irreversible, financial or external actions.	Approvers may lack context or become fatigued.
Audit and rollback	Reconstruction and recovery.	Preserve prompts, tool calls, identities, diffs and deployment versions.	External side effects may not be reversible.

Who should use which part

User or team	Primary requirement	Best starting surface	Likely production route
Prompt or API learner	Test a Gemini model and obtain working code.	AI Studio.	Gemini API in an application.
Individual software developer	Edit, test and understand a local repository.	Antigravity, Gemini CLI or Code Assist.	Existing software delivery system.
Maintainer with queued GitHub work	Delegate bounded tasks asynchronously.	Jules.	Reviewed pull request and normal CI.
Application team	Build a customer-facing agent or workflow.	ADK plus Gemini API.	Agent Platform, Cloud Run or GKE.
Data team	Give agents governed access to analytical and operational data.	ADK, data tools and MCP connectors.	Google Cloud data services plus managed agent runtime.
Platform engineering	Standardise templates, identity, evaluation and deployment.	Agents CLI and ADK.	Managed runtime, observability and policy gates.
Regulated enterprise	Control data, identity, audit and human approval.	Code Assist Enterprise or governed custom agents.	Agent Platform with verified regional and security controls.

Direction and roadmap signals

ONE HARNESS

Antigravity across surfaces

Google is joining desktop and terminal work around a shared agent harness, while retaining Gemini CLI as an open-source route. The likely direction is greater capability parity, not necessarily one merged product.

FULL-STACK BUILD

AI Studio beyond prompts

AI Studio now generates multi-file React/Node and Android applications, uses an Antigravity-derived harness and connects to GitHub and Cloud Run. This moves it towards application construction without making it the complete enterprise lifecycle.

OPEN CONSTRUCTION

ADK as the portable layer

ADK 2.0 expands languages, graphs, providers, protocols and deployment choices. Google benefits when the agent framework can enter heterogeneous estates and still lead to Google models or Cloud operations.

MANAGED OPERATIONS

Agent Platform as control plane

Runtime, session, memory, evaluation, tracing and governance are converging into a managed platform. This is the commercially durable layer because operating agents is harder than demonstrating them.

PROTOCOLS

MCP plus A2A

Tool interoperability and agent interoperability are becoming distinct platform primitives. Adoption depends on identity, policy, observability and compatible semantics, not protocol support alone.

MANAGED AGENTS

Reusable agents through APIs

Google is exposing more pre-built and managed agent capability alongside general models. The boundary to watch is how much orchestration remains transparent and portable.

Roadmap discipline. These are directional inferences from current product releases and documentation, not promises of dates, feature parity or product consolidation.

Limits, risks and unresolved questions

Portfolio complexity. Overlapping surfaces, names and generations raise selection, training and migration costs.
Preview dependence. Agent mode, managed tools or model features can change before general availability.
Authority confusion. Teams may mistake model reasoning, project instructions or a tool schema for permission to execute.
Prompt and tool injection. Files, websites, repositories and connected services can contain instructions designed to redirect an agent.
Credential concentration. A useful production agent can aggregate access across code, data and business systems.
State ambiguity. Local, hosted and managed surfaces do not share identical persistence, tenancy or deletion semantics.
Evaluation gaps. Passing a demonstration or benchmark does not establish reliability across changing real workloads.
Vendor coupling. ADK and open protocols reduce framework lock-in, while managed memory, evaluation and operations can deepen platform dependence.
Cost and latency. Multi-agent loops, large context, retries, tools and managed services multiply consumption even when each model call appears efficient.
Observability fragmentation. Local tools, IDEs, hosted workers and cloud agents can produce different logs and audit trails.
External side effects. Code, browser and connected-system actions can create consequences that a model cannot fully inspect or reverse.
Documentation drift. Command sets, product boundaries and service names are changing quickly; live product documentation remains authoritative.

Primary sources

Google I/O 2026 developer highlightsAntigravity 2.0, AI Studio, managed agents and developer-platform direction.
Build apps in Google AI StudioAgent harness, web and Android build modes, integrations and deployment.
Antigravity overviewDesktop command centre, agents, tools, browser work and review artifacts.
Gemini Enterprise Agent PlatformManaged runtime, sessions, memory, evaluation, tools, observability and governance.
Full-stack development in AI StudioReact, Node, secrets, GitHub and Cloud Run boundaries.
Antigravity CLI overviewTerminal surface, shared harness and multi-step development workflow.
Antigravity CLI featuresCurrent core slash commands and interface capabilities.
Antigravity CLI permissionsDeny, ask and allow rules, precedence and sandbox distinction.
Antigravity pluginsSkills, agents, rules, MCP servers and hooks as extension bundles.
Antigravity MCPConnection and tool exposure through Model Context Protocol.
Gemini Code Assist overviewIDE, lifecycle and enterprise product boundaries.
Gemini Code Assist agent modePreview status, planning, tool loop, permissions and context.
Introducing Gemini CLIOpen-source terminal agent, tools and launch scope.
Gemini CLI documentationInstallation, authentication, configuration and workflows.
Gemini CLI commandsCurrent CLI, slash and at-command references.
Gemini CLI toolsBuilt-in file, shell, web and memory tools.
GEMINI.md context filesHierarchical project instruction and memory import behaviour.
Jules documentationAsynchronous coding-agent concepts and workflow.
Jules tasks and repositoriesVM isolation, repository access, plans, logs and diffs.
Jules is now availableGeneral-availability milestone and product role.
Agent Development KitADK 2.0 scope, languages, model choice, workflows and deployment.
ADK agentsAgent composition from model, instructions and tools.
ADK sessions, state and memoryInteraction identity, working data and cross-session retrieval.
Agents CLI referenceAgent-project lifecycle commands and coding-agent integration.
Agents CLI quickstartScaffold, playground and deployment workflow.
Agents CLI evaluationEvaluation generation, grading and improvement cycle.
Agents CLI deploymentInfrastructure and managed deployment routes.
Agent Builder documentationGoogle Cloud's build, scale and governance services for production agents.
Gemini API function callingTool schemas, model proposals and application-owned execution.
Gemini API toolsBuilt-in and developer-defined tool capabilities.
Gemini Computer UseGraphical-interface action loop and safety boundary.
A2A: agent interoperabilityProtocol purpose, task exchange and ecosystem rationale.
Developer's guide to agent protocolsFunctional distinction between MCP and A2A.
Gemini CLI extensionsPackaged prompts, MCP integrations and distribution model.

Research cut. Facts were re-verified to 24 August 2026, 20:45 ICT. Sources are first-party Google, Google Cloud and project documentation. Product descriptions and maturity labels establish Google's published position; they do not independently prove quality, reliability, adoption or economics.

Part II — Technical reference. These operator controls are grouped by product surface. They support the architecture above; they are not interchangeable and they do not grant execution authority by themselves.

Command appendix

The three command layers below solve different problems. Antigravity slash commands control an interactive agent session; Agents CLI commands manage an ADK project's lifecycle; Gemini CLI retains a separate open-source command system. Custom skills, plugins and extensions can add dynamic commands and therefore cannot be exhaustively listed as native controls.

Antigravity CLI — current core slash commands

Command	Purpose	Boundary
/resume or /switch	Resume or change the active conversation.	Changes session context, not workspace state.
/rewind or /undo	Return to an earlier conversation point.	Does not guarantee reversal of external side effects.
/rename <name>	Rename the current conversation.	Presentation metadata only.
/permissions	Inspect or manage command permission rules.	Deny takes precedence over ask, then allow.
/model	Select an available model.	Entitlement and surface determine choices.
/keybindings	Inspect or configure keyboard bindings.	Local interface behaviour.
/statusline	Configure terminal status information.	Display control, not task state.
/tasks	View or manage background tasks.	Task availability depends on the current harness.
/skills	Inspect installed skills.	Skills can add dynamic invocations and instructions.
/mcp	Inspect MCP servers and tools.	Server identity and downstream permissions remain separate.
/open <path>	Open a file or directory in the configured editor.	Path must be within the accessible environment.
/diff	Review workspace changes.	A diff is evidence, not proof that tests pass.
/usage	Show current usage information.	Account and product availability affect reporting.
/logout	End the authenticated CLI session.	Does not revoke all external credentials.

Antigravity management and extension panels

Control	Purpose	Status boundary
/agents	Inspect or manage agent definitions where the surface exposes them.	Documented outside the core CLI command table.
/hooks	Inspect lifecycle hooks supplied through configuration or plugins.	Hook support and commands depend on the installed product version.
/config	Open relevant Antigravity configuration controls where supported.	Not treated here as part of the 14-command core set.

Agents CLI — ADK project lifecycle

Command	Purpose	Stage
agents-cli setup	Install supported agent-building skills and prepare the environment.	Tooling setup
agents-cli create <name>	Create a new ADK agent project.	Scaffold
agents-cli playground	Run and inspect the agent interactively.	Develop
agents-cli eval generate	Generate evaluation cases from interactions or requirements.	Evaluate
agents-cli eval grade	Run graders and produce evaluation results.	Evaluate
agents-cli scaffold enhance .	Apply guided improvements to an existing agent project.	Improve
agents-cli infra	Prepare deployment infrastructure.	Provision
agents-cli deploy	Build and deploy the selected agent target.	Release
agents-cli publish	Publish supported agent artifacts or definitions.	Distribute
agents-cli cmd-info	Show authoritative detail for a command and its options.	Reference

Gemini CLI — high-value native controls

Control	Purpose	Category
gemini	Start an interactive Gemini CLI session.	Entry point
gemini -p	Run a non-interactive prompt for scripting.	Automation
gemini -i	Start interactively with an initial prompt.	Entry point
gemini -r	Resume a previous session.	Session
gemini update	Update the installed CLI where supported.	Maintenance
gemini extensions	Manage extension packages.	Extension
gemini mcp	Manage MCP server configuration.	Integration
/about	Show CLI and environment information.	Information
/agents	Inspect available subagents or agent controls.	Orchestration
/commands	List available commands, including extension-provided entries.	Discovery
/help	Open help for the active version.	Discovery
/init	Create initial project context guidance.	Context
/mcp	Inspect configured MCP servers and tools.	Integration
/memory	Inspect or refresh hierarchical GEMINI.md context.	Context
/model	Select or inspect the current model.	Model
/skills	Inspect available skills where supported.	Extension
/tools	List the tools exposed to the current session.	Execution
/quit	Exit the interactive session.	Session

Dynamic boundary · checked 24 August 2026. Commands vary by product version, operating system, authentication, organisation policy and installed extensions. Use /commands, built-in help and the official command reference as the authority for the active installation.