🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

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

1. Anthropic's Public S-1: What the Numbers Say About Agent Economics Anthropic publicly filed its S-1 on October 26, reporting $16.286 billion in revenue for the first half of 2026 against a $693 million operating loss. The much larger headline net loss includes a fair-value charge, so it must not be read as cash burn — and the filing does not break out how much revenue comes from agents specifically. The money is real; "agent revenue" remains an inference, not a disclosed segment.

2. GPT-6 Cyber: Stronger Security Agents, Tighter Boundaries OpenAI released GPT-6 Cyber, a GPT-6 Astra variant built for authorized security work, scoring 48.5% on ExploitGym versus 32.1% for its predecessor and newly rated Critical under OpenAI's own preparedness framework. The interesting part is scope adherence: it was trained to stay inside an authorized perimeter. A perfect score in one controlled test is not a production safety guarantee, so the operational question becomes whether verified findings justify the orchestration, review and remediation cost.

3. CrowdStrike's Fiddler and Dreadnode Deals: The New Control Layer for AI Agents CrowdStrike agreed to acquire Fiddler AI (agent observability, to ship as Falcon Observe) and Dreadnode (automated AI security testing), with reported prices near $495 million — terms CrowdStrike has not disclosed. This is the clearest signal yet that securing agents is becoming a platform market rather than a feature. Neither deal has closed and the announced products are not yet generally available.

4. Microsoft Agensh: What 1,024 Agents Actually Prove About Scaling Microsoft open-sourced Agensh (MIT), which drops the central orchestrator in favour of agents that claim subtasks, share findings and verify each other's work asynchronously, using Gitea and Mattermost as shared infrastructure. Across five hard ProgramBench tasks the mean test-pass rate rose from 19.31% with one agent to 28.78% with 128. The headline 1,024-agent run covers a single task (Pandoc) and says nothing about cost per successful task — read it as an architecture probe, not a scaling law.

5. GDPx: Why the Leading AI Agent Scores Just 12% on Hard Professional Tasks SemiAnalysis launched GDPx: 206 tasks across 11 deliberately difficult domains, where Claude Opus 5.5 leads at 12.0%, ahead of Gemini 4 Carbon (10.9%) and GPT-6 Astra (9.6%), at average costs of $51.41, $16.29 and $53.99 per task. These scores are not a success rate for everyday office work — they are a stress test of agent-plus-tool systems. The practical takeaway: measure reliability and boundary compliance separately from task completion.

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