Chat and Dispatch

Use the Cloud Agents chat surface, Ask/Plan/Agent modes, and the New Agent Run flow to create review and execute work with precise repository scope.

The Chat tab is where intent becomes work. It is not only a text box. It is a scoped dispatch surface with repository selection, branch selection, agent selection, execution mode, and run-launch actions.

The Chat Header Controls

At the top of the page, the chat header scopes the session before the prompt is sent.

Control	What it does	Why it matters
Repository selector	Chooses which repository the conversation is operating against.	Prevents cross-repo ambiguity.
Branch selector	Chooses the branch context for analysis or change execution.	Keeps plans tied to actual repository state.
Agent selector	Chooses the system or custom agent for the task.	Matches the work to the right operating instructions and model.
Mode selector	Switches between Ask, Plan, and Agent.	Separates guidance, planning, and real execution.
Execute Plan	Runs the currently visible plan as an action.	Useful after validating a plan block in chat.
Execute Latest Plan Block	Executes the most recent plan output.	Keeps the workflow moving without recreating the plan.

Ask, Plan, and Agent Modes

Ask

Use Ask when you need explanation, diagnosis, or repo-aware advice without starting a run that changes code or generates a formal execution path.

Plan

Use Plan when you want the agent to decompose the task, identify likely files, and present the intended sequence before anything executes.

Agent

Use Agent when the task is clear enough to become a real run. This is the execution mode for review and change workflows.

Recommended mode selection

Situation	Best mode
You need a repository explanation or risk assessment	Ask
You want a reversible checkpoint before work starts	Plan
You already trust the scope and want output now	Agent

New Review vs New Execute

Cloud Agents exposes two fast paths in the workspace header and in the new-run modal:

Run type	Primary use	Typical output
Code Review	Analyze an existing change, pull request, or repository delta	Findings, risks, suggested fixes, and review context
Execute Change	Prepare and run a scoped implementation or remediation task	Code change artifacts, validation output, and possible approval requests

Use Code Review when the code already exists and the question is whether it should proceed. Use Execute Change when the question is how to make the change safely.

The New Agent Run Flow

The New Agent Run modal is split into steps. In the current interface the first step is Setup, followed by scope and detail stages.

Setup

The setup step captures the high-level execution shape:

Run type: Code Review or Execute Change
Agent: the system or custom agent that should handle the task
Mode: Ask, Plan, or Agent
Repository: the repository the run will operate against

The modal also surfaces quota context such as runs today, which helps operators avoid starting a run that the workspace cannot complete under current limits.

Scope

Scope narrows the task to the repository state that matters. Depending on the rollout and the repository configuration, this can include branch, pull request context, or related execution boundaries.

Details

The details step is where the task prompt becomes precise enough to act on.

Strong prompts usually include:

the goal
the files or module area
non-goals
validation expectations
approval expectations when relevant

Example execute prompt:

Implement a loading state for the approvals page. Limit changes to the approvals route and tests. Keep the current visual style. Do not alter repository rules or billing code.

Example review prompt:

Review the current pull request for auth and redirect regressions. Focus on unsafe navigation, missing guards, and test coverage gaps. Ignore formatting-only issues.

Chat Timeline and Repository Context

Chat messages can include intermediate thinking markers, execution steps, completed-step markers, and repository context blocks. When a pull request or branch is inspected, the conversation can surface repository objects directly inside the thread so the operator does not need to cross-reference another tool.

This is useful for:

checking whether the agent found the right pull request
seeing the plan before execution
tracing why the agent chose a certain repository path

Dispatch Patterns That Work Well

Review-first pattern

Start in Plan mode.
Confirm the repository and branch.
Run Code Review.
Open the run record in Runs.
Approve or reject follow-on actions.

Controlled execute pattern

Start in Plan mode for the first few tasks in a repository.
Move to Agent mode only after the plan looks correct.
Keep protected paths active.
Require approvals on merge-capable outputs.

Specialist-agent pattern

Choose a custom agent for a narrow task class.
Keep the prompt shorter because the agent definition already carries role-specific instructions.
Use repository rules to keep the specialist inside its intended surface area.

Common Failure Modes

Most avoidable Cloud Agents failures are scope failures, not model failures: wrong repository, wrong branch, wrong agent, or a prompt that leaves too much room for interpretation.

Avoid these patterns:

asking for broad repo-wide cleanup without path boundaries
using Agent mode before checking that the plan matches the task
dispatching against the default branch when the work belongs on a feature branch
mixing review objectives and implementation objectives in one prompt