Models and Agent Systems
When should better agent performance come from the model, the system or its experience?
Start with the first report → See all reports ↓Overview
A better harness can change what an agent achieves with the same model. That raises the first decision in this series: where does scaffolding make a smaller model sufficient, and where do the task’s demands still exceed it? The answer opens two directed branches: allocating work between models and deciding which part of a failing agent to change.
The repair branch then asks what would count as finishing the user’s task. Once the result and its evidence are clear, the final report asks whether retaining that experience changes genuinely later work. The progression moves from capability and intervention to acceptance and reuse, while the model-routing branch remains available as its own continuation of the opening question.
Follow a capability question into interventions, completion and reuse
Locate the capability boundary
Ask when the harness can make a smaller model sufficient.
Choose the intervention
Allocate model effort or test a targeted model–system repair.
Establish the result, then reuse it
Connect completion evidence to lessons for genuinely later work.
The opening report branches into routing and targeted repair. The repair line continues into completion and experience.
Reading order
A recommended path through the enquiry. Each report explains its part; continuation links preserve the original branches.
- Locate the model–system boundaryCan a Harness Make a Small Model Match a Large One?
Establishes the opening question: when can scaffolding help a smaller model reach the required result, and where do task limits remain?
Start here - Allocate the modelLarge Model, Small Model, or Both?
Follows one original branch into per-request model selection, prediction quality, escalation and the full cost of the routing decision.
- Choose the repairWhen an Agent Fails, Should You Change the Model or the System?
Returns to the other branch: diagnose a concrete failure and compare a stronger model with a targeted system change.
- Establish what done meansWhen can an AI agent say the task is complete?
Makes completion a judgment about the requested result and the evidence supporting it, including when the agent should continue or seek help.
- Turn outcomes into reusable guidanceWhen an AI agent’s past becomes useful
Asks when retained episodes and conditional methods improve genuinely later work rather than merely supporting another retry.
Further reading
A focused look at whether procedural skills add useful capability to an existing agent harness.
Develops the attribution question behind learning from experience: which earlier choice actually changed an outcome?