Headless Harness · Research orchestration prototype
Make parallel AI research a bounded job.
A Python orchestrator decomposes tasks, runs isolated workers and assembles structured results.
Product evidence
See the working surface.
research-job-0423 workers · running
WORKER 01Market evidence
Scanning primary sources
WORKER 02Technical review
7 sources · structured output
WORKER 03Counter-case
Checking unsupported claims
OUTPUT CONTRACT
result.json · citations.json · failures.jsonIsolated workspacesThe problem
Parallel agent research needs explicit boundaries for work, failure and output. This implementation makes those controls part of a job contract.
Choices in the implementation
- Use separate worker directories and explicit timeouts.
- Coordinate independent agent processes in Python rather than relying on an unreliable nested-spawn path.
- Collect structured worker results before optional synthesis.
What exists
- Task decomposition, concurrent subprocess execution, timeout/error handling and typed result artifacts.
- Local execution plus Docker packaging and a Kubernetes job template.
The next question
Compare sequential and parallel runs on the same research tasks, including cost, latency, citation quality and partial failures.