Agents Are Not Enough
AI “agents” — systems that autonomously act on a user’s behalf, often powered by LLMs and tools — are attracting hype but also skepticism over vague definitions, unreliable behavior, and security risks like prompt injection. Commenters debate whether most real value today comes not from fully autonomous agents but from more modest patterns such as “LLM as UI,” low-code workflow generation, or single-call tool use with humans firmly in the loop. Many argue that agency should be treated as a continuum and that economic pressures, reliability constraints, and clearer technical taxonomies will ultimately determine where and how agent-like systems are actually deployed.
LLMs as UI and Workflow Builders
- Several comments propose using “LLM as UI”: the human remains the true agent, the LLM is just a front-end to tools/CRUD APIs.
- A popular variant: use LLMs to generate workflows/macros from natural language, then save and run them deterministically without the LLM.
- Advocates say this reduces UI complexity, helps onboarding/discoverability, and lets non-experts automate complex apps (e.g., CRM/ERP).
Determinism, DSLs, and Verification
- Skeptics worry about non-determinism: even at temperature 0, parallelism and floating-point issues can produce variation.
- Critics argue that if users must verify workflows, they effectively need to understand an underlying DSL; in that case, a GUI or direct DSL may be simpler.
- Others counter that workflows can be summarized in natural language and tested on sample data; raw DSL exposure might only be needed for power users.
- Some note that natural language is imprecise; attempts at precision tend to turn into unreadable “legalese.”
Definitions and Hype Around “Agents”
- The thread notes that “agent” has long been poorly defined and now covers everything from thermostats to LLM-plus-tools.
- Multiple commenters say the term is blurry or becoming meaningless, similar to “data science,” and is heavily driven by marketing and funding.
- There is disagreement over whether “agents” should mean any acting program, or only systems with goals, internal world models, and autonomy.
Practical Value and Current Limits of LLM Agents
- Some see agents as “LLM calls in a loop” with low real-world success rates, expensive token usage, and error compounding.
- Others think agents will improve as tool-calling and parameter mapping get better, but note current multi-turn accuracy is still weak.
- A cited view (from another source) is that many tasks are better solved with single LLM calls plus retrieval, not full agents.
Safety and Ecosystem Concerns
- Prompt injection and secure tool use are seen as unsolved problems for powerful autonomous agents.
- There is speculation that ecosystems will naturally settle different levels of human vs. machine agency based on economic incentives.
Reception of the Paper Itself
- Several commenters find the paper vague, high-level, or “academic theater,” especially its cognitive architecture section.
- Others still find value in its attempt to frame and critique current “agent” narratives, even if underspecified.