Building Effective "Agents"
Anthropic’s guidance on “building effective agents” argues that today’s practical value from large language models comes from simple, well-defined workflows rather than fully autonomous agents, and that heavyweight orchestration frameworks often add more complexity than benefit. Commenters share experiences that favor deterministic code plus narrow LLM calls over open-ended agents, emphasizing reduced state spaces, testing, durable execution, and human-in-the-loop designs to keep systems reliable in production. There is also active debate over what should count as an “agent” at all—legally, philosophically, and architecturally—as vendors market increasingly autonomous-sounding AI systems whose behavior remains fundamentally probabilistic and error-prone.
Workflows vs “Agents” in Practice
- Many commenters agree the real current value is in workflow automation rather than fully autonomous agents.
- “Agentic” planning steps are often removed in production because companies want predictable, repeatable behavior; what’s left is a deterministic workflow with LLM calls.
- DAG-like or state-machine workflows with clear decision trees (e.g., email triage, customer support, marketing flows) are seen as much more deployable than open-ended agents.
- Narrow, well-defined tasks with constrained choice spaces work far better than “boil the ocean” general-purpose agents.
Frameworks vs Simple Patterns
- Strong sentiment against heavyweight frameworks like LangChain/LangGraph for many real-world cases: seen as adding verbosity, indirection, and complexity without proportional benefit.
- Several practitioners report ripping out such frameworks and replacing them with plain code and simple async patterns, often with large LOC reductions.
- Lightweight abstractions or “transitional software design” is favored: mostly traditional code, with LLMs used only for “LLM-hard” subproblems (creative writing, fuzzy classification, nuanced decisions).
Reliability, Testing, and Limits of LLMs
- Recurrent theme: LLMs are inherently probabilistic and unreliable; you must test hard, with archives, edge cases, and regression suites.
- Some argue occasional failures are acceptable in domains already dominated by probabilistic methods (e.g., spam-like tasks); others warn that “magic thinking” about reliability is widespread and dangerous.
- Concerns about compounding errors in multi-step workflows; mitigation via verifiable checks, tests, and possibly LLM-as-judge.
- Complaints about lack of robust structured output and degradation of JSON-like patterns, making agents brittle.
Definitions and Ethics of “Agents”
- Long debate over the term “agent”:
- One camp emphasizes the legal/ordinary sense (acting on behalf of someone, with responsibility).
- Another cites the AI-textbook sense (anything that perceives and acts in an environment).
- Some see current marketing use of “agent” as confusing or disingenuous for non-technical audiences, especially when responsibility for errors is at stake.
Architecture, Orchestration, and Tooling
- Several view LLMs as components in larger orchestrated systems: prefrontal-cortex analogies, system‑1 vs system‑2 splits, and complex multi-model architectures are discussed.
- Durable workflow engines (Temporal, Inngest, Hatchet, Windmill) are highlighted for long-running, fault-tolerant, human-in-the-loop processes, though they introduce their own complexity and learning costs.
- Tool use is seen as central to powerful agentic workflows, but defining tools and handling stateful tools correctly is non-trivial.