Claude's multi-agent API enables defining an "orchestrator" agent to manage "delegate" agents, each with unique toolsets. This creates a programmable, specialized team that mirrors human organizational structures, providing a sophisticated model for tackling complex, multi-faceted problems programmatically.

The "Outcomes" feature requires a markdown "rubric" to define success. This forces developers to codify what "done" looks like, allowing the AI agent to self-grade and iterate up to 20 times. This introduces a structured, testable approach to achieving reliable results from agentic systems.

Anthropic's new "Agent Teams" feature moves beyond the single-agent paradigm by enabling users to deploy multiple AIs that work in parallel, share findings, and challenge each other. This represents a new way of working with AI, focusing on the orchestration and coordination of AI teams rather than just prompting a single model.

AI agents have become proficient at following a pre-defined strategy to execute tasks. The next major frontier, and a significant bottleneck, is the ability to explore open-ended environments and generate novel strategies independently. This is the core capability that benchmarks like ARC AGI v3 are designed to test.

The effectiveness of agent loops lies in their ability to spin up specialized sub-agents. A common framework involves a 'planning agent' that outlines steps and an 'evaluating agent' that quality-checks the output. This division of labor allows the AI system to tackle complex tasks more reliably than a single agent could.

The path to improving production agents isn't manual analysis but automation via other agents. The vision is for every deployed agent to have a "nurse agent" companion. This trainer constantly analyzes production traces, runs experiments by replaying scenarios with different models or tools, and automatically optimizes the primary agent.

The next layer of abstraction involves "strategies" or "meta-harnesses" where tokens are treated as non-fungible resources assigned specific jobs like "advising," "executing," "reflecting," or "grading." This enables more sophisticated agent orchestration and better cost/performance tradeoffs.

To improve the quality and accuracy of an AI agent's output, spawn multiple sub-agents with competing or adversarial roles. For example, a code review agent finds bugs, while several "auditor" agents check for false positives, resulting in a more reliable final analysis.

OpenAI identifies agent evaluation as a key challenge. While they can currently grade an entire task's trace, the real difficulty lies in evaluating and optimizing the individual steps within a long, complex agentic workflow. This is a work-in-progress area critical for building reliable, production-grade agents.