ADR-003: Model Routing and Executor-Native Subagents

Status: accepted

Praetor needs cost-aware model selection, but should not become a first-party

subagent orchestration framework in v1.

Decision

Praetor will support model routing through provider-neutral model profiles:

• fast
• standard
• reasoning
• critical

Each provider or executor maps those profiles to concrete model names or native

settings. Praetor mission planning and agent assignment should reference the

profile, not hard-code provider-specific model names.

When complexity is uncertain, Praetor should choose conservatively and route to

reasoning. Work involving security, privacy, legal, finance, credentials,

runtime configuration, destructive actions, or final high-impact review should

route to critical unless the owner has explicitly configured otherwise.

Praetor will not implement first-party subagent lifecycle management in v1.

For development and research tasks, Praetor may allow executor-native subagents

when the selected executor supports them, such as Codex or Claude Code.

Praetor owns:

• mission scope
• agent assignment
• model profile
• executor selection
• permissions
• budget
• audit trail
• result review

The executor owns:

• internal native subagent usage
• internal parallelization
• internal context isolation
• executor-specific model mechanics

Praetor receives one normalized executor result.

Policy

Minimum runtime policy fields:

model_profile: fast | standard | reasoning | critical
native_subagents: disallowed | allowed

Optional later fields:

native_subagent_policy: none | review_only | executor_managed
max_parallelism: integer

Default behavior:

• simple status, formatting, and summarization work: fast or standard
• ordinary implementation or analysis: standard
• uncertain, cross-file, architectural, or repeatedly failing work: reasoning
• high-risk or final governance review: critical
• native subagents: disabled for simple work, allowed for selected development

or research executor runs

Praetor should not expose a subagent management UI in v1. If native subagents

are allowed, the UI should present this as an executor behavior setting, not as

company staff.

Result Contract

When an executor uses native subagents or dynamic workflows, the normalized

result should include:

• whether native subagents were used
• broad purpose summary, such as research, implementation, review, or tests
• changed files
• commands and tests run
• final result
• risk notes
• recommended human review

Praetor does not require every internal subagent turn. Raw executor logs may be

kept as evidence when available, but the mission view should show the normalized

result and reviewable evidence.

Consequences

This keeps Praetor focused on company governance while still benefiting from

native executor capabilities as Codex and Claude Code improve.

Tradeoffs:

• Praetor cannot perfectly inspect every internal subagent decision.
• Result quality depends on executor result normalization.
• Cost and parallelism must be bounded by Praetor policy before the run starts.

This is preferable to adding Praetor-owned subagent orchestration before the

mission, executor, memory, and review systems are fully stable.