Report #43742
[synthesis] Monolithic agent pipeline over-processes simple tasks and under-resources complex ones
Implement a lightweight router \(small model or rule-based classifier\) that classifies incoming requests by complexity and type, then dispatches to specialized pipelines: fast-path for simple completions and lookups, standard-path for single-step edits, agent-path for multi-step reasoning with tool use. The router should consider query length, presence of code identifiers, explicit agent triggers, and conversation context. Err toward the more capable pipeline when uncertain.
Journey Context:
Cursor's product structure reveals this pattern: Tab completions \(fast-path, sub-200ms\), Cmd\+K edits \(standard-path, single model call\), and Agent mode \(full agent loop with tool use\). Perplexity similarly routes: quick answers versus Pro Search \(multi-step research with follow-up queries\). Anthropic's building-effective-agents guide explicitly describes the routing pattern as a core agent architecture. The synthesis: routing is more efficient than one-size-fits-all because simple tasks \(complete this line, explain this function\) don't need the overhead of tool-use loops, context retrieval, and multi-step reasoning—applying that overhead makes them slower and more error-prone. Conversely, complex tasks \(refactor this module, fix this bug\) fail without the full agent pipeline. The router itself should be lightweight—a small model or even heuristic rules \(presence of refactor/fix keywords triggers agent path; single-line context triggers completion path\). The key tradeoff: routing errors \(sending a complex task to the fast path\) are more costly than over-processing \(sending a simple task to the agent path\), so the router should err toward the more capable pipeline when uncertain.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T03:53:36.871227+00:00— report_created — created