Research agent

For agents that read 100+ sources, plan, draft, and iterate over a single output. Quality matters more than per-token cost. Sessions can run for hours.

What this pipeline is good at

Sessions that span thousands of turns without losing structure (ipc).
Persistent learning across research projects (patterns).
Optionally: pulling earlier-archived sources back when relevant (rehydrator).

The pipeline

Env var

PRXY_PIPE='ipc,patterns'

YAML config

pipeline:
  - ipc:
      targetUtilization: 0.85   # use most of the context window
      preserveLastTurns: 10
      archiveToBlob: true       # keep originals for rehydration
  - patterns:
      maxInjected: 8            # inject many — research benefits from priors
      minSuccessRate: 0.6
      forgeOnSuccess: true

Optional production additions:

PRXY_PIPE='ipc,patterns,rehydrator,router'

Why this order

ipc first — measures and compresses before any other module touches the prompt.
patterns — injects priors after compression, on top of the (now smaller) prompt.
rehydrator — pulls archived turns back when the user references them.
router — picks the model based on the type of step (planning vs drafting vs verifying).

Why no caching?

Research is inherently exploratory — semantically-similar prompts often deserve different answers (different angles, different framing). Caching would lock the agent into yesterday’s interpretation.

If you’re building a research chatbot rather than a research agent, add semantic-cache with a high threshold (0.95+) at the front of the pipeline.

Cost vs quality dial

Use case	Adjust
Quality is paramount (research papers, legal analysis)	Use Opus exclusively. Add `cost-guard` with a generous monthly cap. Skip `router`.
Volume + quality (research at scale)	Wait for `router`. It’ll route easy steps to Sonnet and hard steps to Opus.
Prototype (just exploring)	Add `cost-guard` with `perRequest: 0.30` to keep accidental Opus blowouts bounded.

Variants

Long-form drafting (essays, reports, codebases):

pipeline:
  - ipc:
      targetUtilization: 0.9   # squeeze more context
      preserveLastTurns: 15    # keep recent drafting context verbatim
  - patterns:
      maxInjected: 10
      minSuccessRate: 0.5

Multi-source research with attribution:

pipeline:
  - ipc:
      targetUtilization: 0.8
      archiveToBlob: true       # critical — sources can be rehydrated
  - patterns: { maxInjected: 5 }