From prompt to result
Prepare session context
The session engine builds or refreshes its system prompt with the selected mode, workspace, permitted tools and project rules, then adds the user's message and any supported image attachments.
Request a model response
The configured LLM receives conversation context and the tool definitions available in the selected mode. Context summarization may be used when configured.
Run permitted tools
If the model returns tool calls, the application runs the corresponding local tools and records results, including error results. The tool boundary applies workspace path policy.
Continue or stop
The loop continues until a final response, cancellation, an error or the configured step limit. Activity and execution detail show application-recorded steps and tool I/O.
What execution detail shows
The application help describes Tools & Logs and Execution Detail for prompts/context, model responses and tool arguments/results represented by the app. The visible trace explains application activity; it does not reveal hidden reasoning and does not promise exact replay or a sanitized export.
Mode boundaries
Ask exposes read-oriented tools. Plan can save plan files but does not expose arbitrary writes or shell. Agent enables mutations and shell. By default, writes and shell working directories are constrained to the workspace; a per-session option permits outside-workspace modifications. This is application-level policy, not an OS-level sandbox.
Future multi-agent ideas
Specialist-agent collaboration, execution-plan tracking, workspace knowledge/RAG and The Brain are marked as future features with no backend in the reviewed help. They are not shown here as an operational workflow. See planned work.
Source citations: `uAgentEngine.pas`, `uAgentRunThread.pas`, `uChatSessionHub.pas`, `uAgentExecutionLog.pas`, and `help/how-the-agent-works.md`, with additional permission evidence listed in the feature inventory. Documentation topics →