New techniques for self-evolving, production-grade AI agents

What changed: Requesty AI highlighted five new techniques for making production AI agents more reliable and affordable, including self-evolving agents that automatically refine their own prompts and tools based on real-world feedback, managed multi-agent orchestration layers, and compiled agent workflows that convert flexible plans into more deterministic, cacheable routines. The roundup pulls together research and product announcements from the week of May 19–23, 2026, focused on getting agents out of demos and into mission-critical workloads.

Why it matters: Many teams struggle with agents that are powerful in theory but fragile, slow, or expensive in production; these techniques are aimed at tightening feedback loops, reducing hallucinations, and keeping latency and costs under control. Self-evolving and compiled workflows can cut down on repeated prompt-engineering cycles by letting the system adjust itself under guardrails, while orchestrated multi-agent patterns make it easier to separate planning, execution, and verification steps. Together, they point toward a more “software-engineering-native” way of building agents, closer to how teams ship microservices today.

Try/watch: Audit one high-impact agent use case—like support automation or lead qualification—and map where self-evolving prompts, a verifier agent, or a compiled workflow could reduce retries, API calls, or manual review time; then prototype with explicit metrics for success and a rollback plan.

AI manager “Mona” runs a real-world café in Stockholm

What changed: Andon Labs has deployed an AI agent named Mona, based on Google’s Gemini 3.1 Pro, to autonomously run a café in Stockholm, including hiring and managing human baristas. The deployment is framed as part of a broader trend toward “AI-run companies,” where an agent is responsible for day-to-day operational decisions rather than just assisting a human operator.

Why it matters: This moves agentic AI from back-office workflow automation into visible, customer-facing operations, where the agent is accountable for staffing, scheduling, and service quality. For founders and operators, it’s an early proof point that a single, well-scoped agent can coordinate multiple real-world processes—recruiting, rota planning, and task allocation—under clear constraints. It also surfaces practical questions about labor relations, compliance, and risk management when an AI system is effectively acting as a line manager.

Try/watch: If you run a physical or hybrid business, start with a “shadow AI manager” pattern—let an agent generate hiring shortlists, shift plans, and daily task lists for one team, but keep a human as the formal decision-maker; measure whether the agent’s proposals reduce manager workload without hurting performance or morale.