Neural nets for agents: hardcoding is cheaper than reasoning models
The closed PAC1 benchmark, in which an AI agent must read logs, find files and sidestep prompt injections, unexpectedly exposed a weakness of reasoning models. According to the author, they hallucinated, broke JSON and burned through their API budget by looping on errors. In the end, the task was handled better by Zero-Cost Agent, a deterministic hardcoded executor that simulates a neural network's behavior but runs more stably and cheaply in a formal sandbox.
AI-processed from Habr AI; edited by Hamidun News
The closed PAC1 benchmark for AI agents showed that expensive reasoning models are not always better: a hackathon participant built Zero-Cost Agent from simple algorithms, which beat them on cost and completed the tasks more efficiently.
Why did a simple algorithm beat an AI model on cost?
Rigid algorithms avoid the typical mistakes neural networks make on structured tasks, such as working with logs, files and emails. Neural networks are vulnerable to Indirect Prompt Injections, where a malicious instruction is hidden inside the data. Against such traps, rigid rules work more reliably and cheaply.
Why is hardcoding cheaper than neural networks?
Because it does not require API calls to expensive reasoning models and uses rigid algorithms to perform tasks such as reading logs, finding files and sending emails.
- Oct 28, 2026Anthropic reveals how it builds Claude with Claude: the brain separate from the hands
- Oct 28, 2026Bayer Launches Multi-Agent Engine on AgentCore: 80% of Questions Handled Without Engineers
- Oct 28, 2026Liquid AI releases LFM2.5-24B-A2B: an agent model with 2B active parameters
- Oct 28, 2026AgentOrch: Pre-Planned AI Agents Now Trail Step-by-Step Execution by Just 8.5 Points
Key Questions Answered
Why did a simple algorithm beat an AI model on cost?
Rigid algorithms avoid the typical mistakes neural networks make on structured tasks, such as working with logs, files and emails. Neural networks are vulnerable to Indirect Prompt Injections, where a malicious instruction is hidden inside the data. Against such traps, rigid rules work more reliably and cheaply.
Why is hardcoding cheaper than neural networks?
Because it does not require API calls to expensive reasoning models and uses rigid algorithms to perform tasks such as reading logs, finding files and sending emails.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.