Model Tier Reference

Last Updated: 2026-04-13 Owner: Hone (Skill & Command Architect) Scope: Every agent spawned via the Agent tool

Purpose

Not every agent needs Opus. The structured prompts, agent definitions, skills, and scripts do the heavy reasoning lifting. The model's job is often "follow these instructions" -- which Sonnet and Haiku handle well. This reference classifies every agent by the minimum model tier required to maintain output quality.

Tier Definitions

Opus ($15/$75 per M tokens)

Required when the agent must:

Synthesize across multiple domains, sources, or organizational boundaries
Detect subtle patterns -- enforcement violations, relationship signals, quality drift
Write content with voice and tone -- articles, client emails, identity-specific output
Make architectural or strategic decisions requiring novel judgment
Evaluate quality against subjective criteria (not checklists)

Sonnet ($3/$15 per M tokens -- 5x cheaper than Opus)

Appropriate when the agent:

Gathers data and reports following structured instructions
Classifies or matches against clear, documented criteria
Recommends actions against explicit rules (not novel judgment)
Searches and summarizes from defined data sources
Coordinates tasks following a defined workflow

Haiku ($0.80/$4 per M tokens -- ~19x cheaper than Opus)

Appropriate when the agent:

Executes a specification mechanically -- create record, format output, run script
Performs pure lookups with no interpretation required
Applies templates without creative adaptation
Counts, validates, or checks with binary outcomes
Acts as a write gateway executing validated mutations

Decision Framework

When classifying a new agent, ask these questions in order:

Does it write prose that humans read? (articles, emails, briefs with narrative voice) --> Opus
Does it detect violations or subtle signals? (enforcement, relationship intelligence, quality) --> Opus
Does it synthesize across 3+ data sources into novel output? --> Opus
Does it follow structured instructions to gather and report data? --> Sonnet
Does it classify, match, or recommend against documented criteria? --> Sonnet
Does it execute a specification with no interpretation? --> Haiku

If the answer to #1-3 is yes for any, the agent is Opus. If #4-5, Sonnet. If only #6, Haiku.

Contextual Overrides

Lead vs Teammate

BU Leaders and AI Leaders are Opus when acting as leads (synthesizing agent outputs, making decisions, writing deliverables). When a command spawns them as teammates providing data, they can run at Sonnet if the command's Team table specifies a narrow, structured assignment. The tier listed below assumes the agent's primary operating mode.

V as Session Orchestrator

When V is the primary session model (Sonnet-primary experiment, started Apr 13), V's orchestration role — reading commands, running scripts, spawning agents, presenting results — runs on Sonnet. The structured commands provide the reasoning; the model follows them. V as a spawned subagent (e.g., in /daily-recap review or /meeting-prep architecture assessment) stays Opus because those tasks require cross-domain synthesis.

Build vs Design

Agents with both architectural (Opus) and execution (Sonnet) modes:

Showcase: Opus for architectural decisions, novel interactive UX, template design. Sonnet for routine rebuilds from existing templates (extract data, build, deploy, quality gates). Verified Apr 13: Abs Company portal rebuild on Sonnet — 43 tools, quality gates passed, 5 minutes.

Verified Tiers (Apr 13, 2026)

The following tiers have been verified in production with no quality degradation:

Ledger (Haiku): 4 successful executions — recap listing, Abs Company persistence with full read-back verification (25s), 5P listing, session log (35s). Write gateway on Haiku is proven.
Sentinel (Sonnet): Abs Company engagement check — correct health score, cadence, trend, risk assessment.
Marshal (Sonnet): Abs Company tasks/projects — correct open tasks, deliverables, project stage. Also: 5P project creation with 16 tasks (119s).
Pavilion (Sonnet): Abs Company portal freshness — correct staleness detection, session delta, walkthrough identification.

Complete Agent Tier Assignments (88 Agents)

AI Leaders -- Always Opus

Agent	Codename	Tier	Rationale
V	v	Opus	COO. Synthesizes all agent outputs, writes briefings with voice, makes operational decisions. As session orchestrator: Sonnet (see Contextual Overrides).
Sage	sage	Opus	CCO. Relationship intelligence, writes with emotional awareness, detects subtle signals.
Pax	pax	Opus	CFO. Financial synthesis, judgment on commercial health, writes with analytical voice.
Oracle	oracle	Opus	Strategic advisor. Methodology interpretation, framework-grounded guidance, novel judgment.

BU Leaders -- Opus (as leads)

Agent	Codename	Tier	Rationale
Marquee	marquee	Opus	Media BU leader. Content production decisions, cross-domain media synthesis.
Provost	provost	Opus	Academy BU leader. Education strategy, curriculum decisions, learning path design.
Trellis	trellis	Opus	Collective BU leader. Membership strategy, partnership judgment, community health.
Foundry	foundry	Opus	Apps BU leader. Product strategy, developer experience, quality gate decisions.
Exchange	exchange	Opus	Store BU leader. Commerce strategy, cross-BU transaction design.
Relay	relay	Opus	Service Delivery BU leader. Engagement quality judgment, capacity allocation decisions.

Customer Intelligence -- Mixed

Agent	Codename	Tier	Rationale
Sentinel	sentinel	Sonnet	Structured data gathering: session gaps, engagement changes, declining patterns. Matches against defined thresholds.
Herald	herald	Opus	Session preparation requires synthesizing relationship history, signals, and context into a narrative brief.
Scribe	scribe	Opus	Transcript processing requires synthesis across raw conversation, methodology detection, and intelligence extraction with judgment.
Tide	tide	Sonnet	Interest pipeline tracking: structured queries, stage movement detection, threshold matching.
Scout	scout	Sonnet	New business signals: pattern matching against defined progression criteria.
Quorum	quorum	Sonnet	Office Hours intelligence: attendance counting, question depth classification against defined scale.
Horizon	horizon	Sonnet	Renewal tracking: date math, timeline proximity, structured risk scoring.
Counsel	counsel	Opus	Deal conversation assessment requires subtle judgment across Four Conversations framework.
Primer	primer	Opus	Pre-call coaching card synthesis: distills strategy + relationship context into actionable working memory.
Framer	framer	Opus	Three-option proposal architecture: strategic skeleton design requires creative judgment.
Analyst	analyst	Sonnet	Post-call review: compares transcript against pre-call strategy using structured rubric.

Content and Communication

Agent	Codename	Tier	Rationale
Forge	forge	Opus	Content multiplication requires creative transformation of insights into multiple formats with voice.
Baldwin	baldwin	Opus	Journalist. Writes rich narrative articles weaving methodology, episodes, and shows. Creative writing.
Correspondent	correspondent	Opus	Client email drafting requires relationship-aware tone, voice calibration, and contextual sensitivity.
Broadcast	broadcast	Sonnet	Distribution tracking: structured post scheduling, engagement data gathering, channel management.
Catalog	catalog	Sonnet	Content indexing: classification, metadata extraction, census maintenance against schema.
Vault	vault	Sonnet	Content DB operations: FTS5 queries, structured data retrieval, census reporting.
Chorus	chorus	Sonnet	YouTube comment ingestion and sentiment classification against defined categories.

Operations

Agent	Codename	Tier	Rationale
Marshal	marshal	Sonnet	Task tracking: structured HubSpot queries, milestone status checks, workload distribution reporting.
Beacon	beacon	Sonnet	Status reporting: git log compilation, file change detection, structured output formatting.
Echo	echo	Opus	Pattern memory requires cross-incident synthesis, novel pattern recognition, and learning extraction.
Mender	mender	Sonnet	Error resolution follows defined diagnostic procedures and fix patterns.
Klaxon	klaxon	Sonnet	Alert processing: threshold evaluation, queue management, escalation against defined rules.
Hone	hone	Sonnet	Cross-layer validation: path resolution, count comparison, reference checking. Structured analysis.
Loom	loom	Haiku	Background worker orchestration: script execution, schedule management. Mechanical.

HubSpot and Integration / Platform

Agent	Codename	Tier	Rationale
Ledger	ledger	Haiku	Write gateway. Executes validated HubSpot mutations. No interpretation -- validates then writes.
Canon	canon	Haiku	Write gateway. Executes validated Sanity mutations. Schema validation then write.
Navigator	navigator	Sonnet	HubSpot implementation PM: progress tracking, structured reporting against project plan.
Pavilion	pavilion	Sonnet	Portal activation: structured configuration, staleness detection against defined thresholds.
Sync	sync	Sonnet	Context synchronization: structured comparison between local files and HubSpot records.
Audit	audit	Sonnet	Data integrity: structured drift detection comparing two data sources.

Documents and Knowledge

Agent	Codename	Tier	Rationale
Architect	architect	Opus	Document creation: scoping briefs, implementation specs require judgment and synthesis.
Archivist	archivist	Sonnet	Memory curation: staleness detection, organization against defined criteria.
Scrivener	scrivener	Sonnet	Documentation maintenance: updating references, fixing stale content following patterns.

Media

Agent	Codename	Tier	Rationale
Director	director	Opus	Media production: canvas compositing decisions, guest management, live streaming judgment.
Prelude	prelude	Sonnet	Pre-production checklist: Mux readiness, guest brief lookup, episode status queries.
Encore	encore	Sonnet	Post-production checklist: recording verification, transcript linking, structured pipeline checks.
Curator	curator	Sonnet	Asset management: organizing recordings, tracking media workflows against defined process.
Pixel	pixel	Sonnet	Image generation: follows creative briefs using defined brand guidelines.
Slate	slate	Haiku	Episode pre-creation: mechanical Sanity document generation from schedule data.
Splice	splice	Haiku	Recording linking: connects Mux assets to Sanity documents. Mechanical matching.
Caption	caption	Sonnet	Transcription pipeline: YouTube-to-Gemini orchestration with structured process.
Dub	dub	Haiku	YouTube-to-Mux migration: batch download and upload. Mechanical file operations.
Vigil	vigil	Sonnet	Pipeline monitoring: stall detection against time thresholds, status checking.

Analytics and Intelligence

Agent	Codename	Tier	Rationale
Pulse	pulse	Sonnet	Health scoring: structured metric queries, trend calculation, threshold-based alerting.
Atlas	atlas	Sonnet	Capacity analysis: load calculation, commitment modeling against defined parameters.
Forecast	forecast	Opus	Revenue projections require scenario judgment, what-if modeling with novel assumptions.
Prism	prism	Opus	Contributor intelligence: coaching context requires nuanced human development judgment.
Tuner	tuner	Opus	Enforcement A/B testing: detecting subtle behavior deltas requires calibrated judgment.
Lookout	lookout	Sonnet	SEO scans, broken link detection: structured health checks against defined criteria.
Waypoint	waypoint	Sonnet	CTA analysis: effectiveness measurement against defined engagement metrics.
Nexus	nexus	Opus	Cross-org convergence: synthesizes across V, Sage, Pax domains. Novel coordination judgment.
Tally	tally	Sonnet	Revenue reconciliation: structured comparison of Orders vs Investments vs actuals.
Tributary	tributary	Sonnet	BU revenue aggregation: structured cross-BU data collection and summation.
Mint	mint	Sonnet	Investment monitoring: completeness checking, ROI calculation, spend trending against defined metrics.

Client Operations

Agent	Codename	Tier	Rationale
Settler	settler	Sonnet	Client onboarding: follows defined setup checklist (config files, HubSpot records, portal).
Steward	steward	Sonnet	Contributor onboarding: follows defined onboarding process.
Meridian	meridian	Sonnet	Assessment-to-Interest bridge: structured conversion following defined rules.

Organizational Enablement

Agent	Codename	Tier	Rationale
Aegis	aegis	Sonnet	Org chart health: count comparison, data consistency checks, structured verification. Narrowed from Haiku because website page generation requires some composition.
Q	q	Opus	Quality system management: auditing requires subtle pattern detection, enforcement evaluation.
Dean	dean	Sonnet	Education schema integrity: structured validation of education pipeline data against defined schemas.

Self-Improvement

Agent	Codename	Tier	Rationale
Squire	squire	Sonnet	Code maintenance: dependency updates, cleanup following defined patterns.

Platform

Agent	Codename	Tier	Rationale
Mirror	mirror	Sonnet	Visual QA: screenshot comparison, issue classification against defined quality criteria.
Showcase	showcase	Opus	Portal builder: architectural decisions, novel interactive UX design, React SPA creation. Sonnet for routine rebuilds (see Contextual Overrides).
Link	link	Sonnet	LinkedIn operations: structured message drafting, profile lookups, engagement data gathering.
Meeting Manager (Chronicle)	meeting-manager	Opus	Real-time meeting facilitation: parallel documentation, multi-leader synthesis, live judgment.

Value Path Concierges (12 agents)

All concierges follow a defined journey template with stage-specific rules. They detect signals and guide visitors through the Value Path -- this requires pattern recognition against defined criteria but not novel synthesis.

Agent	Codename	Tier	Rationale
Patron	patron	Sonnet	Sponsor journey: stage-specific guidance following defined Value Path rules.
Envoy	envoy	Sonnet	Consulting client journey: stage-specific guidance following defined rules.
Spark	spark	Sonnet	Catalyst participant journey: onboarding following defined program structure.
Tether	tether	Sonnet	Collective member profile lifecycle: structured application and onboarding process.
Vestibule	vestibule	Sonnet	Office Hours attendee journey: structured engagement guidance.
Usher	usher	Sonnet	Show guest journey: structured guest experience management.
Sylvan	sylvan	Sonnet	Learner journey: structured education path guidance.
Courier	courier	Sonnet	Subscriber journey: structured newsletter/subscription engagement.
Litmus	litmus	Sonnet	Assessment-taker journey: structured assessment guidance.
Leaf	leaf	Sonnet	Article reader journey: structured content engagement guidance.
Alms	alms	Sonnet	Donor journey: structured donation/support guidance.
Accord	accord	Sonnet	Partner journey: structured partnership engagement guidance.

Summary by Tier

Tier	Count	Agents
Opus	28	V, Sage, Pax, Oracle, Marquee, Provost, Trellis, Foundry, Exchange, Relay, Herald, Scribe, Counsel, Primer, Framer, Forge, Baldwin, Correspondent, Echo, Architect, Director, Forecast, Prism, Tuner, Nexus, Q, Showcase, Meeting Manager
Sonnet	54	Sentinel, Tide, Scout, Quorum, Horizon, Analyst, Broadcast, Catalog, Vault, Chorus, Marshal, Beacon, Mender, Klaxon, Hone, Navigator, Pavilion, Sync, Audit, Archivist, Scrivener, Prelude, Encore, Curator, Pixel, Caption, Vigil, Pulse, Atlas, Lookout, Waypoint, Tally, Tributary, Mint, Settler, Steward, Meridian, Aegis, Dean, Squire, Mirror, Link, Patron, Envoy, Spark, Tether, Vestibule, Usher, Sylvan, Courier, Litmus, Leaf, Alms, Accord
Haiku	6	Ledger, Canon, Loom, Slate, Splice, Dub

Cost impact estimate: If all agents ran on Opus, 88 agent-sessions cost 88 units. With tiering: 28 Opus (28 units) + 54 Sonnet (10.8 units at 1/5 cost) + 6 Haiku (0.32 units at 1/19 cost) = ~39 units. Roughly 56% cost reduction with no quality loss.

Borderline Classifications

These agents could reasonably shift one tier in either direction:

Agent	Assigned	Could Be	Why Borderline
Sentinel	Sonnet	Opus	When detecting "emerging concerns" from transcripts, Sentinel does light synthesis. But the criteria are defined thresholds, not novel judgment. Stays Sonnet. Verified Apr 13.
Analyst	Sonnet	Opus	Post-call review involves some qualitative assessment. But it compares against a pre-call strategy card (structured rubric), not open-ended evaluation. Stays Sonnet.
Aegis	Sonnet	Haiku	Mostly count comparison and data checks. But generating website page content requires some composition. Stays Sonnet.
Pixel	Sonnet	Opus	Creative asset generation has a judgment element. But it follows brand guidelines and creative briefs -- not open-ended creative work. Stays Sonnet.
Hone	Sonnet	Opus	Cross-layer validation is primarily structural (path checking, count comparison). This assignment (model tiering) required Opus judgment, but routine Hone work is Sonnet.
Prism	Opus	Sonnet	Coaching context could be structured data gathering. But contributor development requires nuanced human judgment -- genuine "coaching intelligence." Stays Opus.
Echo	Opus	Sonnet	Pattern memory could follow defined incident categories. But novel pattern recognition across incidents requires genuine synthesis. Stays Opus.
Link	Sonnet	Opus	LinkedIn messaging could require voice calibration. But the structured templates and HubSpot persistence workflow keep it in Sonnet territory.
Forecast	Opus	Sonnet	Revenue projections could follow formulas. But what-if scenario modeling requires assumption judgment beyond structured calculation. Stays Opus.

Usage in Commands

When spawning agents via the Agent tool, use the model parameter:

Task(subagent_type: "sentinel", model: "sonnet", ...)
Task(subagent_type: "ledger", model: "haiku", ...)
Task(subagent_type: "forge", model: "opus", ...)

Every command with a Team table should include a Model column specifying the tier for each agent. When Hone audits commands, model tier specification is a checklist item alongside enforcement loading and delegation compliance.

Maintenance

When a new agent is created:

Hone classifies it using the Decision Framework above
The tier is added to this reference
Any commands spawning the new agent get the Model column updated

When an agent's scope changes:

Hone re-evaluates the tier
Updates this reference
Updates all commands that spawn the agent