Field notes
Three kinds of models for agents
"Which model?" is the wrong first question. Production agents combine a reasoning model, a harness that lets it act, and fast decision models that keep it on the rails.
Fast judgment vs deep reasoning
Use the fastest model that can make the call. Save deep reasoning for the steps that need it.
The reasoning engine
Frontier LLMs
Large general models that plan, write, code, and judge. They handle ambiguity better than anything else, but they are slower, pricier, and non-deterministic.
The body around the brain
Agentic harnesses
The runtime that turns a model into an agent: the loop, tools, memory, permissions, and sandbox. The same model can be mediocre in one harness and excellent in another.
Fast, typed judgment
Decision models
"System 1" models that don't write text. You give them a state and a typed question; they return a yes/no, one of your options, or a score, each with a probability. Because the output is constrained to your options, they can't wander off-script.
Frontier LLMs
Pick these for the hard thinking: breaking a goal into steps, handling inputs nobody anticipated, and judging other models' work. Keep providers swappable; the leaderboard changes every few months.
- ✓Planning multi-step work
- ✓Ambiguous or novel tasks
- ✓Writing and reviewing code
- ✓Acting as an eval judge
Rule of thumb
Use a frontier model to plan and to grade, and a faster model for the routine steps in between. That often cuts cost substantially without hurting quality, but let your evals confirm it.
Agentic harnesses
A harness is everything between the model and the world. It decides what the model sees, which tools it can touch, and what happens when it gets something wrong.
- The loop decides, acts, reads results, repeats, with a step limit.
- Tools are exposed through MCP, curated per task.
- Context and memory keep the window small and relevant across long runs.
- Permissions and sandbox limit the blast radius when the agent is wrong.
- Observability records every step so you can replay and evaluate it.
Decision models
A new category in 2026. Instead of generating text, a decision model answers a typed question about a state you give it. They are fast and cheap enough to run on every step of an agent, which makes them a natural fit for the confidence routing and guardrails I describe on the approaches page.
Predicate
“Is this request urgent?”
A probability that a yes/no condition holds. You pick the cutoff.
Choice
“Which team should handle it?”
One of the options you define, with a probability for each.
Score
“How severe is it? (0 low · 1 medium · 2 high)”
A probability-weighted position on your rubric: here 1.1.
Probabilities are estimates, not guarantees. Calibrate thresholds against a set of human-labeled examples before letting them trigger automatic actions. Speed and price figures quoted at launch are vendor claims until you measure them on your own workload.
How they fit together
Decision models guard the doors; the harness and frontier model do the work in between.
| Frontier LLMs | Agentic harnesses | Decision models | |
|---|---|---|---|
| Output | Free text, code, tool calls | Actions and finished work | Typed answer + probability |
| Speed | Seconds to minutes | Minutes to hours | Tens to hundreds of ms |
| Cost per call | Highest | Model cost × steps | Lowest |
| Predictability | Low | Depends on the walls you build | High (fixed answer set) |
| Main risk | Hallucination | Runaway loops and side effects | Uncalibrated probabilities |