| name | parallel-hypothesis-investigation |
| description | Apply when a problem has multiple plausible causes across system layers, or output must pass multiple independent judgment criteria โ to investigate hypotheses or evaluate criteria concurrently rather than sequentially. Use especially when the obvious fix already failed or the problem spans boundaries (config set but behavior wrong). |
Parallel Hypothesis Investigation
When a problem has multiple plausible causes or multiple independent judgment criteria, investigate them concurrently rather than sequentially.
Decomposition Axes
Parallel investigation only pays off when the work decomposes cleanly. Two axes are valid, and they are not interchangeable โ picking the wrong one produces overlapping workers that redo each other's work.
| Axis | When to use | Worker framing |
|---|
| Layer (hypothesis) | Root cause is uncertain; could lie in several distinct layers | "Test whether the cause is in " |
| Criterion (evaluation) | Output must pass multiple independent judgment criteria | "Judge the output against only" |
The layer axis is for finding something; the criterion axis is for judging something. Review and advisor-style work is usually criterion-axis; debugging is usually layer-axis. A single task can use both sequentially โ first find the cause (layer), then judge the fix (criterion) โ but do not mix them in a single parallel batch.
For criterion-axis synthesis output format, see evaluation-integrity.md ยงMulti-advisor synthesis.
When to apply
- The root cause is uncertain and could lie in 3+ distinct areas (library internals, config, infrastructure, etc.)
- The output must satisfy 2+ independent judgment criteria that a single reviewer would blur together
- Each worker can proceed independently without blocking others
- Sequential investigation would waste time if the first guess is wrong
Protocol
- Pick the axis first, then enumerate. Decide whether you are decomposing by layer or by criterion. Then list the workers โ hypotheses for layer-axis, judgment criteria for criterion-axis. Aim for breadth over depth at this stage.
- Dispatch parallel agents. Assign each worker a non-overlapping scope. Each agent should know what to look for (layer-axis) or what to judge against (criterion-axis), and what constitutes a finding.
- Synthesize, don't just aggregate. When results arrive, cross-reference findings. On the layer axis, one agent's discovery may explain or invalidate another's โ the root cause often sits at the intersection of multiple layers. On the criterion axis, agents will often disagree on the same artifact; that disagreement is signal, not noise (see evaluation-integrity.md).
Agent briefing principles
- Give each agent enough context to make judgment calls, not just follow instructions
- Specify the exact directories, files, or functions to start from
- State the hypothesis being tested (layer-axis) or the criterion being judged (criterion-axis), not just "investigate X"
- For criterion-axis agents, explicitly forbid cross-criterion commentary โ a security reviewer should not argue about UX, or the criteria collapse back into one blurred review
- Ask for concise findings, not exhaustive reports
Signals this approach is needed
- You've already tried the obvious fix and it didn't work
- The user says "it should work but doesn't"
- The problem spans boundaries (env vars set correctly but behavior wrong, config exists but not applied)
- Initial investigation rules out the first layer, suggesting the cause is deeper or in a different layer entirely
Rules
- Don't parallelize trivially. If there are only 1โ2 obvious places to check, just check them directly.
- Each agent should cover a non-overlapping investigation scope to avoid redundant work.
- Present synthesized findings as a unified picture with a clear root cause chain, not as disconnected per-agent reports.
- When multiple agents converge on the same finding from different angles, that's strong signal. When they contradict, investigate the gap.
