Building effective agents (Anthropic)

Anthropic’s engineering essay — the canonical first-party definition of what an agent is, and a restraint manifesto the rest of this spoke implicitly argues with.

The workflow / agent distinction

• Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.”
• Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

The split lines up with this spoke’s loop-engineering thread: a workflow is a fixed path; an agent is a loop the model steers itself.

A reliability rationale for the split (tail-control-agentic-workflows). The distinction isn’t only architectural — under a production deadline, the workflow is the reliability-tractable form. Because its steps are fixed and known, they can be raced, cut at p95, and parallelized to control latency’s heavy tail; a self-directing reasoning agent can’t be tail-managed the same way (there’s deliberation to lose). So “orchestrate predefined paths” buys not just predictability of behavior but predictability of completion time — an operational argument for workflows whenever a customer-facing deadline is in play.

The core loop and ground truth

Agents start from a human command or discussion, then “plan and operate independently.” The load-bearing detail: “During execution, it’s crucial for the agents to gain ‘ground truth’ from the environment at each step (such as tool call results or code execution) to assess its progress.” That is the same feedback-signal argument loop-engineering makes — the loop is only as good as the ground truth it can read.

Simplicity and guardrails

Anthropic’s overriding advice is restraint: add complexity “only when it demonstrably improves outcomes,” because “Success in the LLM space isn’t about building the most sophisticated system. It’s about building the right system for your needs.” On bounds, agents “can then pause for human feedback at checkpoints or when encountering blockers,” and it is “common to include stopping conditions (such as a maximum number of iterations) to maintain control” — the first-party root of this spoke’s agent-guardrails thread, and a note of caution against the harness-maximalism elsewhere in the corpus.

Tier T1 — vendor engineering documentation. Cited by agentic-tooling, agent-guardrails, loop-engineering, tail-control-agentic-workflows.