Agent data gets deeper while agent reach gets governed

Curated data that agents can cite

Firecrawl's Alexandria spent the week making specialist corpora askable in plain language. The biggest addition is SEC filings for more than 8,000 public companies, covering 10-Ks, 8-Ks, 13F holdings, and insider trades—the filings that underpin most equity research and much compliance work. In a related move, Firecrawl brought 41 million life sciences papers into ChatGPT and Claude queries, so a question about a drug, disease, or gene returns answers grounded in research literature in a single step rather than a retrieval pipeline you assemble yourself. It also added a free Election Records dataset spanning all 470 House and Senate races, with results as each state reports, campaign funding, and incumbent voting records.

Not everything was a new corpus. Firecrawl benchmarked three search APIs on 581 legal research requests, scoring each against the official source document, and reports its Government & Legal Index leading on both gold@1 and gold@10. It's a useful signal for anyone building retrieval over legal material: the domain-specific index beats generic search at finding the right document first.

Governing where browser agents go

Browserbase shipped Navigation Policies, which let you allow or block whole categories of sites, write rules for specific domains, and reuse a single policy across sessions. You create a policy through the API and pass its ID as browserSettings.riskPolicyId when starting a session; changes apply only to new sessions. The design fits the guardrail shape teams actually need—one auditable policy rather than per-session allow-lists scattered through code—and the category-level blocking is well suited to agent workloads that shouldn't wander onto, say, banking or social sites mid-task. It's opt-in per project for now, enabled via support.

Reliability for Actors

Apify's SDK week was quiet but practical: charged event counts now survive container restarts and Actors emit Event.EXIT on exit, with fixes for reading local Actor input from the storage directory under non-filesystem storage clients, locked shared request-queue retries, request-list URL matching, and Actor.new_client honoring the configured public API URL. For long-running scrapes that get rescheduled, the persisted charge counts matter most—they close a gap where billing state and execution state could drift apart.

Taken together, the week is about agent infrastructure maturing along two axes at once: richer, more citable source material on one side, and tighter control and accounting over what agents actually do on the other.