Pareto Frontier Strategy

5 Review Modes on the Pareto Frontier

Don't pick one magic reviewer. Build an event-driven pipeline that routes each PR to one of 5 non-dominated review modes, with explicit admissibility constraints and an autonomy budget.

Trade-off Dimensions

Each mode represents a non-dominated point on the Pareto frontier

Cost per PR

Token usage and API calls

Latency

Time to first comment

Quality

Actionability and accuracy

Coverage

Issue detection rate

Pareto Frontier Visualization

Cost vs Quality trade-off (non-dominated solutions)

Quality / Coverage

Cost

Cheapest

Balanced

High Quality

Comprehensive

Async

The Five Modes

Each optimized for different trade-offs on the Pareto frontier

Mode A

Single-Shot Structured Reviewer

One Responses call with Structured Outputs. No tools. Fastest and cheapest for routine PRs.

Best for: 80% of routine PRs

Mode B

Risk-Routed Deep Review

Cheap triage for all, deep review when risk score crosses threshold. Best default frontier point.

Best for: Default production mode

Mode C

Tool-First Verifier

Linters, tests, repo queries via function calling. LLM interprets evidence and suggests diffs.

Best for: Quality-critical codebases

Mode D

Specialist Ensemble

Multi-agent: security, perf, API, style agents in parallel. Aggregator merges with strict schemas.

Best for: High-value repos, incident periods

Mode E

Async Bulk Reviewer

Batch API / flex processing for nightly sweeps, regression scanning, second opinions.

Best for: Backlog, continuous audit

Portfolio Recommendation

Keep a portfolio on the Pareto frontier, not one winner

Steady State

Default production portfolio

Mode AMode BMode C

High Value

Critical repos and incidents

Mode D

Background

Backlog and audit

Mode E

Shifting the Frontier

Stepping stones that improve multiple objectives simultaneously

Prompt Caching + Responses Reasoning Items

Repeated prefixes (tools/schemas/system prompts) cut latency and input cost. Cache utilization improves materially with Responses API.

Cost down, Latency down

Better Retrieval (File Search / RAG)

Fewer tokens wasted on irrelevant context, fewer wrong assumptions, better precision on line ranges.

Cost down, Quality up

Cheap-then-verify Pipeline

Small model proposes, large model verifies. "Verify" step lowers error cost without paying for bigger models everywhere.

Cost down, Quality maintained

FPF Patterns Applied

E.13 Problem Frame + proxy-audit (Goodhart control) and E.16 RoC-Autonomy Budget & Enforcement

Agentic Scale

Pareto Frontier Strategy

5 Review Modes on the Pareto Frontier

Don't pick one magic reviewer. Build an event-driven pipeline that routes each PR to one of 5 non-dominated review modes, with explicit admissibility constraints and an autonomy budget.