Clay's New Agents Remember Your Accounts — Even When Claygent Didn't

Clay’s new Account Research Agents don’t just research an account once — they keep a running account history that compounds over time, a capability standard Claygents never had.

Built on top of Clay’s Audiences unified data layer, the agents pull from Gong call transcripts, CRM records, email history, third-party enrichment data, and buying signals, then write synthesized findings directly back into a company’s CRM or data warehouse. The key differentiator from Clay’s existing Claygents is persistence: standard Claygents restart context on every run, while Account Research Agents maintain compounding context, building a running account history instead of a one-off snapshot. The feature ships with a Data Hub dashboard for monitoring agent activity, catching errors, and tracking spend, and highlighted use cases include flagging product-qualified accounts for expansion, resurfacing closed-lost deals when circumstances change, monitoring churn-risk signals, and giving new reps full relationship history at onboarding. It’s live now in open beta across Clay’s Enterprise, Growth, and Launch plans, with no row limits and no manual CRM configuration required.

This is Clay compounding two moves it already made this month. On July 16, Clay opened Claygent to open-weight models Kimi K2.6 and GLM 5.2 specifically because the newest generation could finally sustain the multi-step planning and browsing agentic research needs — not just content generation — which is the reasoning capability persistent account research depends on. On July 14, Clay shipped Clay MCP to general availability, letting ops-governed functions run inside the six AI tools reps already use — Claude, ChatGPT, Claude Code, Codex, Copilot, and Glean — without reps improvising their own tool mix. Put together, the sequence reads as one build: cheaper agentic reasoning, then governed rep-facing distribution, then persistent, compounding account memory.

The top 2% won’t win by turning this on and letting it run. The differentiator is whether someone actually reviews what an always-on agent surfaces about a stalling or reopened account and multithreads that insight into the deal — rather than letting a dashboard full of flagged accounts sit unread.