Frontier AI labs operate behind thick glass walls. Nobody outside knows how close models are to writing their own successor code or how much computing power actually goes toward safety versus raw capability growth. Anthropic tried to crack open that door by publishing three specific internal metrics designed to track the actual velocity of AI research and development.
If you want to understand where the technology is heading without listening to marketing fluff, these three measurement pillars provide the honest baseline.
The Automation Index: Measuring How Much AI Builds AI
The first and most scrutinized metric tracks recursive automation—specifically, the extent to which artificial intelligence builds the next generation of itself. Anthropic maps this using an automation scale running from AL0 (zero AI involvement) to AL5 (fully autonomous operation without humans).
Within this framework, level three means the AI collaborates by handling large chunks of work under close human direction, while level four means the model leads the task end-to-end from a high-level prompt while humans supervise.
The numbers reveal a sharp climb. Back in early 2026, almost zero internal research tasks reached the collaboration threshold. By August of that same year, models operated at or above the collaboration level for more than ninety percent of internal R&D work, with Claude leading twenty-six percent of tasks directly.
Nobody has hit full autonomy yet, but the steady migration of heavy engineering workloads from human hands to automated agents changes the baseline of software velocity entirely.
Agent Oversight: Monitoring Thousands of Active Workers
Tracking the pace of progress requires watching what those agents actually do once deployed. Anthropic monitors roughly 30,000 active AI agents running concurrently on its internal research platforms.
Managing thousands of autonomous workers requires strict guardrails. The system splits oversight into two distinct operational layers:
- Online monitors: Real-time filters that inspect actions before execution, blocking dangerous behavior like unauthorized data exfiltration within seconds.
- Offline monitors: Post-hoc reviews that analyze agent activity logs to spot subtle signs of drift or misalignment over long horizons.
Out of billions of automated decisions processed in a single month, safety filters blocked roughly 0.002 percent, or about one in every 47,000 actions. Most blocks catch formatting errors or misapplied parameters, but the infrastructure proves that scaling up agent workforces demands automated police work just to keep systems stable.
Compute Allocation: Where Safety Budgets Actually Go
The final metric examines energy and resource distribution. Public debates about artificial intelligence often fixate on raw cluster sizes and training runs, but the real question is how much compute protects society versus how much pushes boundaries forward.
Anthropic measures the fraction of its total research compute dedicated exclusively to safety work, alignment, interpretability, and evaluation. In conservative tallies, safety workloads capture a distinct slice of total infrastructure. If a project yields dual benefits for both capability and safety, it gets excluded from the safety tally to keep numbers strictly honest.
Transparent resource tracking matters because industry-wide pacing efforts depend on verifying that labs don't abandon safety work when competitive pressure spikes. Without transparent breakdowns of compute spend, claims of responsible scaling remain impossible to verify.
Tracking these three vectors—automation depth, agent oversight density, and compute distribution—gives developers and policymakers a practical lens on actual industry velocity. Watch the automation index move up the scale, because that shift tells you more about the future of technology than any press release ever will.
