Claude 4.8 and scheduled agents. AI takes action.

Anthropic ships at breakneck pace: Opus 4.8 only 6 weeks after Opus 4.7. Official focus: long-horizon agentic work. On the Super-Agent benchmark, Opus 4.8 is the only model that completes every use case end-to-end — beating prior Opus and GPT-5.5 at cost parity.

Concretely, Claude 4.8 stays connected to a task for hours, keeps context coherent, and finishes what it started. That's the difference between an assistant that completes and an employee that executes.

The Code with Claude Tokyo announcement eclipsed everything else: Claude Managed Agents now have scheduled deployments. Concretely, an agent can be configured to start every day at 9am, open a session, execute its task, commit its result. No external scheduler, no infra to maintain.

Coupled with environment variables stored in vault, it's a paradigm shift: we leave « LLM-as-a-service » for « LLM-as-an-employee ». Our team is already exploring this model for this site's news carousel.

Announced at the same time: Claude Code moves dynamic workflows to General Availability. Concretely, we can now script complex decisional sequences (« if X then Y else Z ») natively in Claude, without third-party frameworks like LangChain or CrewAI. The model becomes orchestrator.

While Tokyo celebrated scheduled agents, Anthropic fired another shot on June 9: the public release of Claude Mythos via Claude Fable 5. Mythos is a cybersecurity-specialized model previously reserved for 50 curated partners (including the US government) through Project Glasswing. Preview tally: over 10,000 critical vulnerabilities detected in strategic infrastructure software.

With Claude Fable 5, that capability finally exits the enterprise ghetto. The model is available to paying customers with reinforced safeguards (blocking offensive cybersecurity and bioterrorism use cases). Every dev team can now audit their own code at a level that didn't exist 6 months ago.

For the apps we build, the cost/value ratio of AI flips. A task that used to require a full-time dev can now run autonomously via a scheduled agent. And with Fable 5, every app we ship now goes through a pre-deployment AI security audit: vulnerability detection, fix suggestions, automated pentesting. A level previously reserved for big tech.

The question is no longer « can AI do it? » but « is the work worth agentifying, and has it passed the Fable 5 audit? ».