🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. Gemini Deep Think Is Crushing Reasoning Benchmarks — and It's Not Just More Compute Google's Deep Think mode now leads ARC-AGI-2 (84.6%), Codeforces, and IMO 2025 all at once — a sweep the benchmark wars haven't seen in months. The edge comes from generating multiple parallel chains of thought at inference time rather than a bigger base model, a signal that the next reasoning gains are architectural and available to anyone willing to pay the compute bill.

2. OpenAI Is Building a 'Persistent' Codex That Works Until You Put It to Sleep Code changes spotted by WIRED reveal a "Persistent mode" that keeps Codex working until explicitly put to sleep, plus a "Proactivity" prompt that generates follow-up tasks across sessions. It's the clearest push yet toward always-on agents — and it lands the same week OpenAI blamed a highly persistent internal model for its Hugging Face hacking incident, putting the alignment risk front and center.

3. Salesforce and Anthropic Launch Claudeforce: The Enterprise Agent Stack Finds Its Default Salesforce and Anthropic made Claude the default reasoning engine across CRM, Slack, and Agentforce, shipping a 37-skill "Salesforce in Claude" plugin. Salesforce raised its FY2027 outlook and the stock jumped roughly 13% after hours — the clearest sign that the enterprise agent race will be won on distribution, data, and governance, not raw model weights.

4. Temporal's State of Development 2026: 80% of Engineers Now Use AI Agents Daily A survey of 554 engineers finds daily agent use jumped from 47.3% to 80.8% in a year, with 92.3% already rebuilding software they used to buy. The blockers have shifted from model quality to state tracking, debugging, and compute costs — the frontier is now infrastructure, not intelligence.

5. Broadcom Ships a Deny-by-Default Runtime for Enterprise AI Agents Broadcom used VMware Explore to launch a "deny-by-default" runtime where every model, tool, and dataset an agent touches must be explicitly bound, enforced by an AgentMinder controller already handling roughly 43M API calls a day internally. Agent governance is becoming its own product category — even as 76% of IT leaders now hold a negative view of Broadcom's VMware ownership.

📖 Read the full coverage: https://the-agent-report.com/latest/ 💌 Subscribe: https://buttondown.com/theagentreport See you next week! — The Agent Report