Skip to content
Agentic Engineering Primer

Capability Patterns

Use this page for durable capability classes, not leaderboard snapshots. Specific rankings move too quickly to build the guide around.

Capability Classes That Matter

Section titled “Capability Classes That Matter”

Deep reasoning models

Section titled “Deep reasoning models”

Best for:

  • multi-file changes
  • architecture decisions
  • debugging with several interacting causes
  • long agent loops that need to recover from failure

Tradeoff:

  • slower and often more expensive in time or usage budget

Fast iteration models

Section titled “Fast iteration models”

Best for:

  • autocomplete and short edits
  • drafting tests or boilerplate
  • quick review loops where latency matters more than depth

Tradeoff:

  • weaker at holding long plans and resolving ambiguous requirements

Multimodal models

Section titled “Multimodal models”

Best for:

  • UI implementation from screenshots or mockups
  • debugging visual regressions
  • working from diagrams, design files, or image-based documentation

Tradeoff:

  • not every multimodal model is equally strong at coding depth

Long-context models

Section titled “Long-context models”

Best for:

  • large investigations
  • broad repository mapping
  • document-heavy workflows

Tradeoff:

  • large context windows help only when context is selective and well-structured

Local or open-weight models

Section titled “Local or open-weight models”

Best for:

  • sensitive code
  • offline or air-gapped environments
  • teams that prioritize control over frontier performance

Tradeoff:

  • capability may lag top hosted models, especially on hard agentic tasks

How to Choose by Workflow

Section titled “How to Choose by Workflow”
WorkflowStart with this capability classWhy
Complex bug fixDeep reasoningRoot-cause analysis matters more than speed
New feature with many moving partsDeep reasoningPlanning and recovery matter
UI build from design referencesMultimodalVisual understanding changes the result
Tight edit loopFast iterationLower latency keeps the workflow moving
Large codebase explorationLong-contextBreadth helps when paired with context hygiene
Sensitive or regulated workLocal or open-weightOperational boundaries may matter more than peak capability

Context Reality

Section titled “Context Reality”

Advertised context is not the same thing as reliable context. Once the prompt gets noisy, even very large windows help less than people expect.

  • prefer selective retrieval over giant prompt dumps
  • treat long context as a tool for breadth, not permission to include everything
  • keep core rules pushed into project context files and retrieve the rest on demand

See Context Engineering for the workflow implications.

What This Page Intentionally Does Not Do

Section titled “What This Page Intentionally Does Not Do”
  • maintain live model rankings
  • promise a single “best model”
  • freeze benchmark snapshots into durable guidance

For time-sensitive benchmark details, use Benchmarks That Matter and confirm current data before making team-level decisions.

Evidence Tags

Section titled “Evidence Tags”
  • Research-backed: verification, selective context, and review costs matter more than leaderboard chasing
  • Practitioner-backed: capability classes are how many teams actually choose models in daily work

The taxonomy on this page is a workflow-first simplification, not one benchmark’s official ontology.

Next Steps

Section titled “Next Steps”