Terraform Deep Researcher

This example builds a deep-research agent in the shape of the modern orchestrator-worker pattern - the same architecture as systems like gpt-researcher. It deploys three roles, each authored as a small project of instruction files: an intake agent that turns a request into a commissioned job, a big-brain orchestrator that runs the research, and a worker that answers a single sub-question with web search and fetch.

The key idea is that the orchestration lives in the orchestrator agent at runtime - the agent decides how to break the work down, fan out, wait, and recombine, rather than following a workflow graph baked into Terraform. The orchestrator decomposes the question into focused sub-questions, creates and runs a worker task for each (they execute in parallel), introspects their status and result summaries, adapts as findings arrive - going deeper on rich threads, pruning dead ends - and then synthesizes everything into one report. The fan-out and the join are ordinary task abilities the agent calls (create, run, list, fetch). Terraform declares only the stable topology: the agents, their tools, the shared workspace, and the entry surface. The research tasks themselves are runtime artifacts the agents spawn on demand.

Results flow through a shared space used as a blackboard - workers write cited findings, the orchestrator reads them back and writes the final report - plus each task's generated summary and full final output. Each bot carries its own model, so the team uses opus to orchestrate and synthesize and a lighter model for the legwork, and per-task time and iteration limits keep "research until satisfied" bounded.

This is the pattern for open-ended research where the plan should adapt to the question and the agents themselves should own how the work is broken down and recombined.