glossary · Running it

Model tier

definition

A model tier is a named level of capability, such as strong, balanced or fast, that an AI agent asks for instead of a specific model, so the workspace decides which model fills each tier and can change it in one place.

in one line: Agents ask for a level; the workspace picks the model.

in context

Where does a model tier fit?

A model for each agent. In ConvOps: the pin for the runtime, else the tier's model, else the executor's own.

pinfor this runtime
tier modelstrong · balanced · fast
executor modelthe last step
The first one that is ready on a connection wins.
two ways to say it

What is a model tier, in plain words?

Same idea at two depths: the plain version, then what the engine actually does.

Hard-coded models go stale. A tier names how much capability an agent needs.

Change the model behind a tier once, and every agent on it runs the new one next.

for engineers

Built-in tiers: strong, balanced, fast, inherit (aliases opus, sonnet, haiku). Custom tiers per workspace.

model_tiers_manage sets one model per runtime. Each run records model and model_text.

the longer answer

Why does a model tier matter?

Agent definitions often hard-code a model. That works until the model is retired, a cheaper one is good enough, or a different runtime needs a different provider. Then every definition has to be edited, and some get missed.

A tier adds one level of indirection. The agent says how much capability it needs; a workspace setting says which model that means today, per runtime. Changing the model behind a tier reaches every agent that uses it on its next run, and an agent that truly needs one specific model can still pin it.

in convops

How does a model tier work in ConvOps?

In ConvOps, an agent names a tier: strong, balanced, fast, inherit, or a custom tier of the workspace (the aliases opus, sonnet and haiku mean the first three). Each tier sets one model per runtime, Claude Code and OpenCode, and inherit means the executor's own model. Every agent runs on its pin for the run's runtime, else its tier's model, else the executor's own model. A pin or tier model counts only once a real call has worked on one of the executor's provider connections, and every run records which model each agent used and why.

in depth

Strong, balanced, fast: what is each tier for?

Model lists are usually read in three bands. Frontier models are the most capable and the most expensive, used where a wrong answer is costly. Large general models handle most everyday work at a fair price. Small models are fast and cheap, and good at narrow, high-volume tasks. Tiers give those bands names an agent can ask for, so the agent describes its job and not a product.

In ConvOps, strong, balanced and fast are built in, together with inherit, and they can never be renamed or deleted. An owner or admin can add custom tiers for the workspace, such as one for long documents. The workspace can also keep an allowlist of the models its tiers and agents may use, and retired models are left out.

The built-in ConvOps tiers, with example agents.
TierUsual bandFits work likeExample agent
strongFrontierJudgment where a mistake is expensiveA contract reviewer that flags risky clauses
balancedLarge generalMost drafting, analysis and codingA writer that drafts the weekly customer update
fastSmallNarrow, high-volume checksA triage agent that labels incoming tickets
inheritWhatever the executor runsAgents with no preferenceA helper that should match the runner it lands on
in depth

Pin, tier or executor model: which wins?

Every run has to settle on one model per agent, and the order matters when the choices disagree. In ConvOps, the order is fixed and recorded, so the answer to "why did this agent run on that model" is always in the run itself.

  1. 1

    The pin

    If the agent pins a model for the run's runtime, and that model is ready on one of the executor's provider connections, the agent runs on it.

  2. 2

    The tier

    Otherwise, the tier's model for that runtime is used, again only when it is ready on one of the executor's connections.

  3. 3

    The executor's own model

    Otherwise, the agent runs on the executor's own model. An agent is never refused because of its model; only a missing or unusable executor model stops a dispatch.

  4. 4

    The record

    The run notes the model each agent used and why, such as "tier balanced" or "executor model: tier inherit", so a fallback is visible instead of silent.

in depth

How do you change a model in one place?

Suppose a new model is released and the team wants every balanced agent on it. Without tiers, someone edits every agent definition, in every repository, and hopes none was missed. With tiers, the change is one edit to the balanced tier, and each agent that uses it picks the new model on its next run. A tier that agents or teams use cannot be deleted, so the change can never leave an agent pointing at nothing.

Each tier sets one model per runtime, Claude Code and OpenCode, so the two runtimes can use different providers under the same tier name. Examples:

  1. Enterprise: a platform team moves every balanced agent across forty workflows to a newer model release in one edit, and the run records show the switch.
  2. Support: the triage agent stays on fast while the reply reviewer runs on strong, and only the reviewer's cost goes up.
  3. Research: an OpenCode runner uses a different provider's model under the same balanced tier as the Claude Code runners.
questions

What do people ask about a model tier?

What is a model tier in ConvOps?

A model tier in ConvOps is a named level of capability that an agent asks for instead of a specific model: strong, balanced, fast, inherit or a custom tier of the workspace. Each tier sets one model per runtime, Claude Code and OpenCode, so the workspace decides which model fills it and can change it in one place for every agent.

Why use a tier instead of naming a model?

A tier lets you change the model behind many agents in one place, and in ConvOps that change reaches each agent that uses the tier on its next run. The agent says how much capability it needs, and the workspace decides which model that means per runtime. A tier that agents or teams use cannot be deleted.

What does inherit mean?

Inherit means the agent has no preference of its own and runs on the executor's own model. In ConvOps it is one of the four built-in tiers, with strong, balanced and fast, and none of them can be renamed or deleted. The run records the reason, such as "executor model: tier inherit", so the choice is visible.

Can one agent ignore its tier?

Yes, with a pin. In ConvOps an agent can pin a model for a runtime, and that pin comes first: pin, then tier, then the executor's own model. A pin or tier model counts only once a real call has worked on one of the executor's provider connections; otherwise the next choice is used and the reason is recorded.

Which model does a ConvOps agent run on?

Its pin for the run's runtime, else its tier's model, else the executor's own model, each counted only when it is ready on one of the executor's provider connections. ConvOps never refuses an agent because of its model; only a missing or unusable executor model stops a dispatch. Every run records which model each agent used and why.

Who chooses the model tiers in a workspace?

Whoever looks after model choice and cost for a workspace, usually a platform lead. ConvOps model tiers let that person change the model behind many agents in one place, so agent authors ask for strong, balanced or fast and never edit each agent when a provider ships a new model.