Tereza Tížková

The Stanford HAI 2024 AI Index reports a 30x productivity gap between AI leaders and laggards. The differentiator is not company culture, prompting technique or model selection, but the infrastructure. Organizations capturing outsized value from AI agents have machine-readable codebases, deterministic internal APIs, CI/CD pipelines with agent-addressable hooks, and permission models granular enough to scope exactly what an agent can touch. I believe the “agents as employees” framing is most useful if you operationalize it. An employee has persistent identity, episodic and semantic memory, scoped permissions that don’t get renegotiated every task, an audit trail, and a defined escalation path when things go wrong. Persistent computer use (with a stable execution environment that survives across steps) was the real inflection point that is making this possible. Some interesting production problems remain under-explored. How do you give an agent persistent identity across pull requests? How do you recover from partial failure mid-task without discarding completed work? How do you enforce code ownership policies when the author is a model? How do you bound token spend when pipelines spin up sub-agents recursively? This talk defines agent readiness as a concrete infrastructure checklist: structured codebases, deterministic APIs, per-agent scoped credentials, atomic and idempotent operations, structured execution traces, and explicit thresholds for when the agent stops and a human takes over. It presents research results in practice, and what are the steps organizations need to take to be fully agent-ready.