Smart Model Routing in Datarim 1.6: Right Model for Every Agent
Datarim 1.6 introduces model assignment — each of the 16 agents and 14 task-skills now runs on the right model for its role. Opus for critical reasoning, Sonnet for standard work, Haiku for structured tasks.
Claude Code's capabilities come with real costs. When every agent runs on Claude Opus, every framework command draws from the same high-cost quota — even agents doing simple, structured work that doesn't require it. Datarim 1.6 fixes this with smart model routing: each agent and skill now runs on the model that matches its actual cognitive requirements.
The Problem with One-Size-Fits-All
In Datarim 1.5, all 16 agents defaulted to Claude Opus. That made sense as a safe starting point, but it was economically wasteful. Consider what each agent actually does:
- A tester agent checks code against a fixed test suite — structured, deterministic, no deep reasoning needed
- A developer agent writes standard feature code following established patterns — solid work, but not critical-path reasoning
- A strategist or architect makes irreversible architectural decisions that affect everything downstream — this genuinely needs Opus
Running a tester on Opus is like hiring a chess grandmaster to sort your email. The capability is there; the application doesn't require it.
The Model Assignment Convention
Datarim 1.6 introduces a clear decision matrix for model selection:
| Model | Use When | Examples |
|---|---|---|
| opus | Critical reasoning, irreversible decisions, high-stakes analysis | architect, strategist, security, skill-creator |
| sonnet | Standard implementation, writing, analysis, most pipeline work | developer, compliance, writer, devops, sre |
| haiku | Simple structured tasks: formatting, linting, test execution | tester |
Reference skills — those that provide checklists, patterns, and standards rather than doing independent reasoning — don't specify a model at all. They inherit from the calling agent, which already has the right model for its context.
How It Works in Practice
Claude Code supports a model: field in agent and skill frontmatter. Datarim now uses this for every agent and task-skill. The assignment is done once at framework level — no per-task configuration needed.
# Example: architect.md frontmatter
---
name: architect
model: opus
description: Chief Architect for system integrity and scalability
---
# Example: developer.md frontmatter
---
name: developer
model: sonnet
description: Senior Developer implementing features with TDD
---
# Example: tester.md frontmatter
---
name: tester
model: haiku
description: Platform QA Tester for functional and integration testing
---The result for the full set of 16 agents:
| Model | Count | Agents |
|---|---|---|
| opus | 6 | architect, planner, strategist, security, reviewer, skill-creator |
| sonnet | 9 | developer, compliance, code-simplifier, devops, editor, librarian, optimizer, sre, writer |
| haiku | 1 | tester |
For skills: 14 task-skills (those that do active work) have model assignments using the same matrix. 6 reference-skills (datarim-system, ai-quality, security, testing, performance, tech-stack) inherit from the calling agent.
Built Into the Framework
Model assignment isn't just about existing agents — it's now part of how you extend Datarim. The skill-creator agent (which generates new skills, agents, and commands) now requires a model: field in every artifact it creates. The /dr-addskill workflow includes an explicit model determination step.
This means every future addition to the framework comes with the right model assignment from day one — it's a convention, not an afterthought.
What Stays on Opus
The decision to keep 6 agents on Opus wasn't arbitrary. Each of these agents makes decisions that are hard or impossible to reverse:
- architect — system design and architectural patterns; wrong calls propagate everywhere
- planner — task decomposition and complexity assessment; bad plans waste implementation work
- strategist — build/don't-build decisions; scope and approach for the whole task
- security — vulnerability identification; a miss here has production consequences
- reviewer — final quality gate before archive; must catch everything the others missed
- skill-creator — generates new framework artifacts; quality here affects every future task
The cost of a bad decision in any of these roles exceeds the savings from using a cheaper model. Sonnet and Haiku are used where the task is well-defined and the cost of error is low — structured work, not judgment calls.
Datarim 1.6
Model assignment is part of the 1.6.0 release, which also includes unified task numbering (project/area-specific IDs that persist through the full task lifecycle) and an expanded skill set to 20 skills.
If you're using Datarim, update from GitHub and run ./install.sh to get the new model assignments across all agents and skills.
git pull origin main
./install.shDatarim is open source (MIT License) — github.com/Arcanada-one/datarim.