How to build an agent harness with Parallel Search
Blog post from Parallel Web Systems
A research-agent harness built around Claude Opus 5.5 and Parallel’s Search and Extract APIs manages the tool-use loop, web retrieval, context window, spending, retries, citation validation, logging, and evaluation so the agent can answer live-web questions with traceable sources while stopping safely at defined limits. The design uses strict JSON-schema tools, defaults searches to the faster and less expensive mode while restricting advanced searches, caps returned excerpts, and relies on server-side context editing to remove older tool results without disrupting Claude’s thinking blocks. A cost ledger tracks model and retrieval charges against per-run budgets, reserves funds for final answers, and pairs step caps with explicit outcomes such as done, uncited, budget, or step limit. Tool failures are returned to the model as errors, transient API failures are retried, and citation checks require factual answers to reference valid source IDs before they can be marked complete. JSON-line traces and a fixed-question evaluation suite support regression testing of accuracy, costs, completion rates, and tool behavior, while a local replay setup enables inexpensive CI testing. The guide also contrasts direct API integration, which offers detailed control over modes, budgets, excerpts, and citation IDs, with Parallel’s MCP server, which simplifies prototypes and Agent SDK workflows but provides less custom control.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.