Bundled agents

The Admin UI and the CLI ship almost the same set of agent definitions (src/main/resources/initial-data/agent-definitions/) — the Admin UI adds one extra, summarizer. They're loaded on first start and you can edit or add to them.

An agent is system prompt + LLM config + tools. All of them reference the agent-default LLM config, an alias that delegates to openai-default out of the box — repoint the alias and every bundled agent follows.

Assistants & orchestrators

Agent	What it does	Key tools
assistant-with-tools	General developer assistant — runs commands, reads/writes files, browses documents, sends email, and calls other agents.	bash, file_*, read_document, run_agent, todo_write, gmail_*
planner	Orchestrator: decomposes a large task into sub-tasks and delegates each to a specialized sub-agent.	run_agent, run_agents, todo_*, workspace_*
research-lead	Research orchestrator: splits a question into sub-topics, dispatches them to web-researcher in parallel, then verifies and writes a report.	run_agents, run_agent, todo_*, workspace_*

See the research-lead flow for research-lead end to end.

Web research sub-agents

Agent	What it does	Key tools
web-researcher	Focused web research; finds pages and delegates reading them to url-reader, returns structured findings. Uses openai-default.	web_search, run_agent
url-reader	Reads one web page (static or JS-rendered) and returns only the relevant excerpts.	web_read, web_read_browser
verifier	Adversarial check: given a claim and its evidence, tries to refute it.	web_read, file_read, read_document, bash

Code & filesystem sub-agents

Agent	What it does	Key tools
explorer	Read-only sweep of the filesystem/codebase to locate files and naming conventions.	glob, file_list, file_read, bash
code-analyst	Read-only deep analysis: reads code and explains how it works.	file_read, glob, file_list, bash
file-finder	Locates files by name, type or location hint.	file_list, glob
document-analyst	Answers questions over long documents (PDF, Word) using outline/grep/page-range tools.	document_outline, grep_document, read_document

Office sub-agents

Agent	What it does	Key tools
presentation-builder	Builds one PowerPoint deck from a brief — or revises one through the spec it saved beside it — in the chat's working directory, and returns only the path and slide titles. default-chat hands it presentations, so the drafting stays out of the chat's context. Ships with the pptx-builder skill.	code_execute, skill pptx-builder

Creative

Agent	What it does
poet	Crafts poems on any topic.

Internal helpers

These are used by the runtime itself (summarization, titling, quality checks) rather than driven directly:

Agent	What it does
conversation-summarizer	Summarizes older turns for working-memory compression.
tool-summarizer	Compresses large tool results after the first LLM turn.
summarizer	Writes concise Markdown summaries of any text it is given (Admin UI only).
title-generator	Generates a short conversation title from the first message.
answer-relevance-checker	Reviews a draft answer for relevance and returns a verdict.

Skills

coding-assistant and default-chat ship with skills switched on and no names selected, which means every skill this installation, the user and the session's project have. It costs nothing while there are none: the skill tool is only added to an agent that actually has one to load, and the prompt's skills section only appears then. The Admin UI and the CLI seed one example skill, pull-request-review, so the screen is not empty on a fresh install.

Customizing

Edit any JSON file under initial-data/agent-definitions/ to change a prompt, swap the llmConfigName, or adjust the tool list. In the Admin UI you can do all of this from the agent screens. Add a new file to register a new agent; reference it from an orchestrator via run_agent("<name>", "<task>").