Faster LLMs, Bigger Demands: Why Coding Agents Won’t Stabilize Soon

LLM coding agents are powerful but currently feel like dialup: unreliable, slow, and prone to flakiness amid surging token demand. Faster throughput (tok/s) could unlock parallel, semi-unsupervised workflows—but semiconductor limits, reliability issues, and pricing pressures will shape how this scales. Developers who stay curious and adapt their tooling will capture the most productivity gains.

Key Points

• Reliability is shaky across major LLM providers, and agentic coding workflows are flakey and resource-hungry, echoing the dialup era.
• Even limited OpenRouter data shows explosive token growth; agentic coding likely uses ~1000x more tokens than basic chat.
• Throughput (tok/s) is now a key UX constraint; at ~2000 tok/s, the human becomes the bottleneck and new parallel, semi-unsupervised workflows become viable.
• Demand will keep compounding as capabilities rise, but semiconductor stagnation limits efficiency, stressing infrastructure and reliability.
• Expect pricing to shift (e.g., off-peak incentives) to flatten demand; developers who adopt and adapt agent tooling will gain the most.

Sentiment

The community is sharply divided. A vocal camp sees LLM coding agents as genuinely transformative, sharing specific productivity wins in bug hunting, boilerplate generation, and overcoming creative inertia. An equally vocal camp views the entire premise skeptically, arguing that perceived productivity gains are illusory, hallucination makes LLMs fundamentally unreliable, and the 'you're using it wrong' defense is unfalsifiable hype. The article's specific thesis about infrastructure instability receives less attention than this underlying disagreement about whether coding agents are valuable enough to create the demand surge the author predicts.

In Agreement

• Speed improvements and new agent patterns create insatiable demand that will keep outpacing infrastructure capacity
• Current agent speeds break developer flow, forcing teams to adopt new multi-session paradigms that are still maturing
• The tooling ecosystem is rapidly fragmenting across Cursor, Claude Code, Codex, and Zed with no stable consensus emerging
• Even enthusiastic users report frequent need to course-correct agents, suggesting the technology hasn't stabilized
• Provider reliability is genuinely concerning as developers become increasingly dependent on external AI infrastructure
• Off-peak pricing and other innovations will be necessary as peak-time congestion worsens

Opposed

• Research suggests AI increases the perception of productivity without actual gains, undermining the premise of growing demand
• Hallucination is a fundamental unsolved limitation that prevents LLMs from being reliable coding tools
• 'You need to know how to use it' is an unfalsifiable defense that mirrors previous tech hype cycles like Agile and XP
• Verifying AI-generated code can be harder than writing it yourself, negating supposed time savings
• LLMs are only useful for boilerplate and throwaway scripts, not for complex work on large codebases
• The cognitive overhead of supervising agents and reviewing their output may exceed the cost of just writing code directly