RESEARCH, WRITING, AND KNOWLEDGE AGENTS

DEEP-RESEARCH AGENTS THAT GATHER AND SYNTHESIZE SOURCES A deep-research agent is an agent that takes a broad question and works on it over many steps, searching for and reading many sources, and then combining what it found into a structured answer. To synthesize means to combine separate pieces of information into a unified whole that says something none of the pieces said alone. A typical deep-research workflow has these stages. Planning: break the broad question into sub-questions. For example, the question should our company adopt a four-day work week might split into sub-questions about productivity evidence, costs, legal issues, and employee experience. Gathering: for each sub-question, search for sources, read them, and take notes. Evaluating: judge each source for reliability, such as who wrote it, how recent it is, and whether it provides evidence. Cross-checking: compare sources against each other and notice where they disagree. Synthesizing: combine the findings into an organized report. Reviewing: check the draft for gaps, errors, and unsupported claims. This workflow uses ideas you already know. The planning stage uses the planning techniques of Part Three. The gathering stage uses tools such as web search and document reading from Part Four, and it may use the retrieval techniques of Chapter 47. The note-taking uses scratchpads and the knowledge habits of Chapter 48, which also protect the work when the context window fills and is compacted, as in Chapter 21. Large research tasks are often split among several agents, with a lead agent assigning sub-questions to helper agents, a pattern from Chapter 43. This can speed the work, but it also multiplies cost, so the economics of Chapter 53 apply. Beware common weaknesses. Source quality: search results can include low-quality, outdated, or deliberately misleading pages, and an agent may not tell the difference. Bias toward the first results: the agent may stop after finding a few agreeable sources. Missed contradictions: sources that disagree may be blended into a single smooth statement that hides the disagreement. Prompt injection: web pages can contain instructions aimed at the agent, as discussed in Chapters 38 and 62. Overconfidence: the final writing may sound certain even when the evidence is thin. Good harnesses counter these with requirements to report disagreements, to state confidence levels, to prefer primary sources, which are the original documents or data rather than someone's summary of them, and to flag gaps honestly.