Generality–Strength Tradeoff

Today's LLMs trade generality (working across any domain) against strength (reliably solving complex tasks); narrowing scope is how you buy the latter.

Classic ML models were narrow specialists (sentiment analysis, fraud detection, translation). GPT-style models perform an enormous variety of tasks across domains, but this is not AGI (artificial general intelligence) — AI meeting or exceeding human-level cognition, able to assimilate knowledge, reason about it, solve novel and complex problems, and generate genuinely new knowledge with humanlike creativity across any domain. Today's LLMs are markedly deficient at reasoning and problem-solving (especially mathematics), mostly recombine existing knowledge rather than introduce new knowledge, and cannot learn new information outside of training.

A future AGI would possess both strength (solving complex problems) and generality (solving problems in any domain). Current LLMs sit on a tradeoff curve between the two:

  • A pure chat app with an open system message sits at the general-but-weak end — it will discuss anything but won't reliably accomplish complex tasks, and is best suited to one- or two-step tasks where a human does the actual work.
  • Narrowing an agent's domain — a custom system message plus domain-specific tools, i.e. system prompt architecture and a scoped tool inventory — trades generality for strength.
  • Pushing further along the strength axis, LLM workflows break a large task into small, well-defined tasks executed with high fidelity, coordinated by a supervisor process that may or may not itself be LLM-based. A workflow doesn't handle arbitrary requests — it's built for one specific task, and is therefore more capable at that task than a general-purpose conversational LLM agent asked to improvise the same multi-step plan.

Conversational agents fail complex workstreams for concrete reasons: they lack natural work-item queues, grow distracted as system message and tools bloat, and treat instructions as suggestions rather than enforceable structure — failures seen when a single agent is asked to crawl, invent, draft, and send marketing at once. This tradeoff is why a task's shape should drive the architecture choice: a one-off, exploratory, or genuinely open-domain request favors a conversational agent; a recurring, well-understood, multi-step task favors narrowing scope, up to and including a full workflow. See conversational agent structural limitations for the specific mechanics of why conversational agents break down on complex multi-step work. Prefer modular workflows under progressive agent architecture; reach for LLM-driven workflow routing only when that strength still needs open-ended agency.

Sources

  • Prompt Engineering for LLMs (Berryman), ch. 9