Ask HN: Is anybody getting value from AI Agents? How so?
Many developers experimenting with AI “agents” — autonomous or semi-autonomous LLM-driven workflows — report that they’re still too slow, unreliable, and fragile for most high‑stakes or complex tasks. The most consistent value so far comes from tightly scoped uses like coding assistance, customer support triage, simple sales outreach, web scraping orchestration, and personal automations, usually with a human firmly in the loop. Overall sentiment is cautiously optimistic about future potential as models improve, but current hype around fully autonomous agents is seen as premature, with significant legal, UX, cost, and trust issues still unresolved.
Overall sentiment on AI agents
- Many commenters report disappointment with current “agent” systems (AutoGPT-like, LangChain agents, OpenAI Assistants, etc.): too slow, expensive, fragile, and unpredictable.
- Strong sense that agents are “too early”: models still make frequent errors, diverge in loops, and need heavy supervision.
- Some think we’re entering a post‑hype correction phase; others are bullish long‑term but bearish on near‑term deployment at scale.
Where agents / LLMs are delivering value
- Coding assistance (not full agents): Copilot, GPT‑4, Cursor, aider, plandex, etc. help with boilerplate, refactoring, tests, parsing compiler errors, and multi‑file edits—always with a human in the loop.
- Customer support / ops: AI assistants handling a large share of chats, email classification, drafting replies, order cancellations, FAQ bots, Discord helpers. Often paired with fallback to humans.
- Internal workflows: summarizing docs/books, RAG chat over papers, security‑audit diffs, sales research on websites, social‑media signal scanning, form filling, intent classification for chatbots.
- Personal tools: custom Telegram or Shortcuts-based assistants tied into calendars, reminders, RSS, 3D printers, news “radio shows,” etc.
- E‑commerce and marketing: personalized cold email generation, site conversion optimization agents reportedly lifting click‑through.
Technical and product challenges
- Reliability: agents choose wrong tools, ignore tools, hallucinate tasks, get stuck in loops, or disbelieve correct tool outputs.
- Latency and cost: chain‑of‑thought and multi‑step plans make systems slow and expensive; p99 latencies of several seconds to tens of seconds.
- Error compounding: multi‑step workflows magnify per‑step error; attempts at error‑checking with more LLM layers have limited success.
- Architecture: complex agent graphs and rule‑based orchestration feel like a regression to brittle, edge‑case‑ridden systems; some advocate small, tightly constrained agents plus deterministic state machines or workflows.
- Models: current LLMs seen as bottleneck; many think better base models will suddenly make existing agent code “just work.”
Trust, legal, and social issues
- Legal risk: liability for AI mistakes (e.g., hallucinated promises), copyright status of AI outputs.
- Low trust: advice to treat AI like a high‑schooler; only use for low‑stakes or easily reviewed tasks.
- Broader concerns: demand may be weaker than investors assume; environmental costs and creative‑industry impacts raised.