Skip to content
v0.1.0MIT

Build and run AI compliance agents by chatting. Describe an agent in plain English and PitCrew drafts it, publishes it to your workspace, and runs it on your documents, without leaving the conversation.

bulk-run-automation

Run an existing PitCrew agent in bulk over N items at once: N PDFs, N rows of a spreadsheet, N files in a Drive folder, N objects under an S3 prefix. Activate when the user asks to process a batch, a folder, a sheet, a mailbox export, or anything phrased as "for each of these, run the agent".

chat-mode

Answer questions about an existing agent — tool choices, data mappings, workflow logic, what a tool does, why a decision was made. Activate when user asks a question and state.agent_spec exists. Read-only — never modifies state.

config-mode

Generate the configurations block (per-run settings panel) for a completed agent. Analyse tool_call and llm_call tasks for hardcoded params that users should be able to tune per-run. Activate after an agent has been built and knowledge has been generated.

edit-mode

Build, validate and test an agent from an approved plan. Activate when state.plan exists and user wants to build the agent, or when user requests changes to an existing agent structure, mappings, or tools.

file-handling

Read and analyze files uploaded by the user. Use when the user references a file, data, document, CSV, or PDF, or when uploaded files are relevant to the current request. Files are at /files/, workspace cache at /workspace/.

knowledge-mode

Generate the knowledge section for a completed agent: good_examples, careful_zones, and known_limitations. Activate after an agent has been built and validated, when knowledge needs to be documented.

loop-mode

Add a loop swimlane to an agent. Activate when the user's workflow needs to process each item in a list — e.g. "for each email", "for every client", "process each row". Loop is a swimlane-level feature, not a task_type.

orchestrator

Route user requests to the correct worker. Activate for every turn. Delegate to plan-worker when user wants to build a new workflow. Delegate to edit-worker when a plan is approved and agent needs building. After edit-worker completes: run knowledge-worker, then config-worker, then present the final agent_spec JSON to the user. Delegate to chat-worker for all questions. Never answer questions directly — always delegate.

plan-mode

Gather workflow intent, discover tools from /tools/ filesystem, clarify requirement gaps, produce a structured plan spec for approval. Activate when user wants to build a new agent, state.plan is None, or user wants to change an existing plan.

studio-build-mode

Step-by-step procedure for building a studio-forms JSON artifact using the studio_* MCP tools. Use when the user asks to build a form, intake, questionnaire, or any client-facing document. Always starts from source material uploaded to a document collection, then covers field types, who fills each field, delivery (email/SMS code, link expiry, reminders) and error recovery.

trigger-automation

Set up a trigger that fires an existing PitCrew agent automatically — on a cron schedule, when an external event happens in Gmail / Drive / Slack / etc., or chained to another agent's bulk run. Activate when the user asks to schedule an agent, run it on an event, or wire two agents together.