The model is the engine. Its capabilities and limitations define what's possible.
Sections
- Model Selection — How to choose the right model for agentic tasks
- Model Behavior — How models behave in agentic contexts—variance, consistency, temperature effects, and behavioral patterns
- Model Limitations and Workarounds — Common model limitations in agentic contexts and practical workarounds—math, hallucination, context limits, instruction drift, and upgrade breakage
- Multi-Model Architectures — When and how to use multiple models in agent systems—orchestrator patterns, cascades, routing strategies, and planning versus execution separation
- Model Evaluation — How to evaluate models for agentic tasks—metrics, benchmarks, observability, and the compound error problem
Core Questions
This chapter explores:
- Selection: How do you choose the right model for a task? What capabilities matter most?
- Cross-provider selection: When does provider choice matter for agentic systems? Which harness ecosystem fits which task type?
- Behavior: What model behaviors help or hinder agentic work? How do you account for variance?
- Limitations: What can't models do? How do you work around constraints?
- Architecture: When do multi-model systems make sense? How do you route between them?
The Short Version
Default to frontier models. The capability gap between SOTA and everything else still matters more than cost optimization in most cases. Downgrade only when you have evidence that a smaller model works reliably for your specific task.
Access tier is distinct from capability tier. As of April 2026, the highest-capability model (Claude Mythos Preview) is not accessible via the standard API — it is restricted to enrolled programs. For standard-API practitioners, Opus 4.6 is the frontier ceiling. See Capability-Gated Access Tiers.
Within the frontier tier, harness ecosystem matters more than model capability differences. Top frontier models (Opus 4.6, GPT-5.2, Gemini 3 Pro) have converged in overall capability. The harness — the autonomous execution environment — is the primary differentiator for agentic task outcomes. Choose provider based on harness availability for your task type before comparing model benchmarks.
Reasoning and tool use are the capabilities that matter most for agentic work. Context length, speed, and cost are secondary — important for architecture decisions, but not for the core question of "can this model do the job?"
Connections
- To Prompt: Instruction-following is shaped more by prompt quality than model selection. Different models respond differently to the same prompt, but SOTA models are generally more forgiving.
- To Context: Context length limits shape architecture decisions. Frontier models tend to have larger windows, but the capability-capacity tradeoff still applies.
- To Tool Use: Some models are better at tool-use than others. This is a core selection criterion for agentic work.
- To Practitioner Toolkit: Harness availability varies by provider. Model selection and harness selection are coupled decisions for agentic systems. See Agent Frameworks for harness capability tiers.