The coordinator’s system prompt shapes how it orchestrates. Step-by-step procedures produce predictable but rigid behavior. Goal-oriented prompts with quality standards produce adaptive behavior that scores 14 points higher on average — because 45% of research queries benefit from strategy changes that rigid procedures can’t make.
The data: procedural vs goal-oriented
| Metric | Procedural prompt | Goal-oriented prompt |
|---|---|---|
| Average quality | 68/100 | 82/100 |
| Adaptation rate | 0% | 45% |
| Quality variance | Low (uniform) | Higher (scales with complexity) |
The 14-point quality improvement comes from the coordinator’s ability to adapt when sub-agent results reveal unexpected findings or gaps. The higher variance in goal-oriented output is healthy — simple queries produce simple reports, complex queries produce complex reports. Forcing uniform output means over-investing on simple queries or under-investing on complex ones.
A/B test: 19-point gap, fixable edge case
500-query A/B test:
- Procedural: 100% completion, 65/100 quality, 35% suboptimal execution
- Goal-oriented: 95% completion, 84/100 quality, 40% adaptive strategy changes
The 5% incompletion is a convergence issue fixable with iteration guards. The 19-point quality gap is fundamental. Procedural “completes” by mechanically following steps — but in 35% of cases those steps were the wrong approach. A complete but wrong report is worse than one that iterates toward quality.
The rigid procedure trap: literal execution
A coordinator prompt says: “Step 1: Use search agent to find 5 papers.” During a quantum computing query, the search agent finds 2 papers and 1 key government report. The coordinator ignores the government report — the prompt says “papers,” not “reports.”
A goal-oriented prompt (“produce comprehensive analysis with diverse sources”) would include the valuable report. The coordinator’s job is to pursue the research goal, not to match a prescribed format for source types.
The balanced approach: goals + optional hints
Neither extreme is optimal:
- Pure goals, no guidance: the coordinator may be aimless for unfamiliar tasks
- Pure procedure, no goals: the coordinator can’t adapt when reality doesn’t match the plan
The effective middle: goals and quality standards as the primary framework, with optional procedural hints for common patterns. “Typically start with broad search before deep analysis” is a hint. “Always start with broad search” is a mandate. Hints bootstrap strategy without constraining it.
Complexity-adaptive coordinator
The same coordinator handles both simple fact lookups and complex multi-domain analyses. The prompt should enable proportional effort:
“Assess query complexity first. For simple factual queries, a single agent pass suffices. For complex multi-domain queries, use multiple agents and iterate until coverage criteria are met.”
This scales naturally: quick resolution for simple queries, thorough investigation for complex ones. One coordinator, adapted strategy.
Strategy vs format: separate the concerns
A CI coordinator needs two things that seem contradictory: adaptive strategy (skip security scan for docs-only PRs) and predictable output (consistent report format for auditors).
The fix: separate them.
- Strategy (which agents, how many passes, what depth) → goal-oriented, adaptive per PR
- Format (report structure, severity ratings, file references) → structured, predictable always
The coordinator adapts HOW it reviews while always producing WHAT auditors expect. Adaptability applies to the investigation process; predictability applies to the output structure.
What to include in coordinator prompts
- Research goals: what to achieve, how broad/deep
- Quality standards: what “good” looks like (cited sources, specific findings, coverage completeness)
- Evaluation criteria: when to iterate vs accept (coverage gaps? depth insufficient?)
- Optional hints: common patterns without mandating them (“typically start with…”)
- Output format requirements: consistent structure regardless of adaptive strategy
- Coordination meta-instructions: “evaluate coverage after each round,” “pass only relevant context to sub-agents,” “preserve both perspectives when sources conflict”
What NOT to include
- Step-by-step procedures as primary structure (prevents adaptation)
- Decision trees for every scenario (unmaintainable, still misses edge cases)
- Minimal instructions without quality standards (no basis for self-evaluation)
- Prescribed agent order without flexibility (“always run security, then performance, then style”)
One-liner: Goal-oriented prompts with quality standards score 82 vs 68 (procedural) because 45% of queries benefit from adaptive strategy — use goals for what to achieve, optional hints for common patterns, and separate adaptive strategy from predictable output format.