When you give Kady a big task, it doesn't have to do everything itself. It can hand parts of the work to sub-agents - independent helpers that run in the background, do one focused job, and report back. Think of Kady as the lead scientist and sub-agents as the lab members it delegates to.
You don't need to do anything to make this happen. Kady decides on its own when delegating makes sense - for example, reviewing code while continuing the analysis, checking every citation in a manuscript, or running three independent analyses at the same time.
Every project comes with 21 scientific specialists pre-installed. Each one is an expert persona with its own instructions and quality standards:
| Area | Specialists |
|---|---|
| Code & computation | code-reviewer, statistical-reviewer, math-checker, ml-auditor, data-validator, reproducibility-auditor, pipeline-engineer, data-visualizer, simulation-reviewer |
| Literature & fact-checking | literature-researcher, citation-checker, fact-checker, methodology-reviewer, peer-reviewer |
| Study design & ideas | hypothesis-generator, experiment-designer, protocol-writer, results-interpreter |
| Writing | manuscript-editor, abstract-writer, ethics-reviewer |
A few examples of what they do:
citation-checkerverifies that every reference in a document actually exists and actually supports the claim it's attached to. Anything it can't verify is flagged as "unverifiable" - never quietly passed.statistical-revieweraudits an analysis for the right test, violated assumptions, sample size, and p-hacking patterns - and re-runs the numbers itself when the data is available.peer-reviewerwrites a full, journal-style referee report on a manuscript: major concerns, minor concerns, questions for the authors, and a recommendation.reproducibility-auditorchecks whether someone else could re-run your analysis from scratch - and actually tries to.
There are also 8 general-purpose agents that ship with the underlying delegation engine (pi-subagents): reviewer, scout, planner, worker, researcher, oracle, delegate, and context-builder.
You can simply name one in your message:
"Use the statistical-reviewer to check the analysis in
results.ipynb."
"Have the citation-checker go through
manuscript.md."
"Run peer-reviewer and methodology-reviewer on my draft in parallel and combine their feedback."
Sub-agents can run one at a time, several in parallel, or chained (one's output feeding the next) - Kady handles the orchestration.
Open Settings (gear icon) → Sub-agents. From there you can:
- See every agent available in the current project, with its description.
- Edit an agent (pencil icon) - change its instructions, give it a different model, restrict its tools, or adjust its thinking depth.
- Add your own agent - click Add agent, give it a name like
assay-qc-checker, write its instructions in plain language, and save. It's immediately available for delegation in new chats. - Delete agents you don't need. Deletions stick - they won't silently come back.
- Restore defaults - brings back the 21 scientific specialists in their original form (your own custom agents are untouched).
- Customize a built-in - the 8 engine agents are read-only, but clicking Customize copies one into your project where your version takes priority.
| Field | What it does |
|---|---|
| Name | How Kady refers to the agent (lowercase, hyphens allowed, e.g. code-reviewer) |
| Description | One line telling Kady when this specialist is the right pick |
| Model (optional) | Make this agent use a specific model - e.g. a cheaper model for routine checks, a stronger one for hard reviews. Leave empty to use sensible defaults |
| Thinking level | How much the agent "thinks before speaking" - higher levels reason more deeply but cost more |
| Tools (optional) | Limit what the agent can do - e.g. read, grep, find, ls makes an agent that can inspect files but never modify them. Empty = full toolset |
| Inherit project context | Whether the agent sees your project's AGENTS.md instructions |
| Inherit skills | Whether the agent can use the project's scientific skills |
| Replace base system prompt | Off (recommended): your instructions are added to the standard agent behavior. On: your instructions completely replace it |
| System prompt | The agent's full instructions - who it is, what standards it applies, and how it should report results |
Each agent is a plain markdown file in your project at sandbox/.pi/agents/<name>.md. The Settings panel is just a friendly editor for these files - you can also view and edit them directly in the file browser. Edits apply to new chat tabs.
Sub-agent work uses your API key like everything else. Their spend is recorded in the same project cost ledger you see in the header, and the project's spend cap applies to them too - once a project hits its limit, Kady is blocked from starting new sub-agents.