What to build instead of AI agents

“Better models will fix it” vs. engineering now

• One camp argues agent frameworks are a stopgap; better models in 1–2 years will make today’s heuristic “LLM call glue” obsolete.
• Others push back: people have said this for years; builders today can’t just wait and will lose in competitive markets.
• Even with stronger models, outputs remain stochastic, so fully autonomous logic without human oversight is viewed as risky.

What counts as an “agent” and as expertise

• Debate over whether people have really been building “agents” for 3–5 years or just scripted LLM calls.
• Some insist agency requires tool use, planning, and multi-step autonomy; simple API calls aren’t agents.
• Broader dispute over expertise: 5 years vs. 10–15 years for “true” mastery in such a fast-moving field.

Plain code and traditional workflows still matter

• Many agree an even more basic point is missing: lots of problems don’t need LLMs at all; “if you can solve it algorithmically, do that.”
• Hype and funding incentives nudge teams to bolt AI onto everything, but most real problems remain simple and deterministic.

Context engineering, memory, and brittleness

• Multiple commenters report that managing context is the main challenge: curating what the agent sees, structuring .md files, and roles.
• Letting agents update their own docs or memory tends to degrade quality over time, requiring human curation.
• This is likened to a return of “feature engineering,” now reborn as “context engineering” due to finite context windows.

Human-in-the-loop, taste, and control

• Several people prefer “tight leash” tools like Claude Code/Cursor: AI writes code or drafts, humans provide taste and direction.
• There’s skepticism that prompts can fully encode personal taste or complex design decisions.
• Trust remains low: agents are useful when you can verify their work faster than doing it yourself.

Agents vs. workflows in automation and enterprise

• Supporters of the article say deterministic business processes and enterprise automation should be hard-coded or orchestrated via workflows, with LLMs as components.
• Critics counter that with top-tier models, natural-language agents can now replace dozens of brittle scripts, especially in messy, evolving domains like incident response.
• Some see agents as expensive “temporary glue” until stable, cheaper non-AI implementations are discovered.

Frameworks, orchestration styles, and future directions

• Several note that many failures come from immature, “toy” agent frameworks and naive coordinator agents.
• Proposed alternatives: declarative control flow, explicit state management, many small focused prompts, and treating agents as functions within workflow/orchestration tools (e.g., Airflow-based SDKs, unified pipelines).
• Others forecast a near-term wave of robust desktop/browser/RPA-style agents, built atop provider SDKs and strong agentic models, further shifting the calculus.

Low-value use cases and scraping

• Spam/sales outreach is criticized as a weak, error-tolerant poster child for agents; simple keyword rules could do the job.
• Web-scraping agents face pushback from infrastructure like Cloudflare; workarounds (vision-equipped browsers, user-side plugins) may remain feasible but more expensive.