Models

A model for each agent.
One team for the work.

Corin proposes models for the agents it creates. Assign a different model to each role, including models from different vendors, then use past work to assess a change.

CorinModel configuration / Example team
Quote review team
FOUR AGENTS · INDEPENDENT MODEL SETTINGS

Choose the model for the role

TEAM / Q-1042
12 MODEL FAMILY EXAMPLES

Choose a candidate model

Intake

Representative model families for evaluation; actual provider and model availability depends on the configured release.

SELECTED AGENTIntakeCurrent: GPT · OpenAI
CANDIDATEQwenAlibaba
Compare token use, latency, cost and reviewed outcomes on representative cases before switching.
Choose any agent, then a model family. Model identifiers and provider access are configured for the release.

Per-agent configuration

Model choice follows the role.

Corin proposes a coordinated team from the work and its operating context. Each agent has its own responsibilities and model setting. A pricing analyst can use one provider while a contract analyst uses another.

You can change one agent’s model without treating the entire team as a single model choice. The handoffs, permissions and relevant memory still belong to the agentic system.

One task, distinct roles

Different models can work on the same task.

In a quote review, an intake agent structures the request, a pricing analyst checks permitted history, and a contract analyst checks the applicable term. Each role can be configured with a different model.

The model is one part of the agent configuration. Its tools, operating knowledge, memory and approval boundaries still shape what it can do.

QUOTE TEAM / Q-1042
WORK ITEMRevised quote request
INTAKERequest structured for the team.

Its model setting belongs to the intake role, not to every agent.

One work item; independent model settings.
MODEL CHANGE / PRICING ANALYST
Should the pricing analyst use Qwen instead of Claude?
CURRENTClaude SonnetAnthropic
CANDIDATEQwenAlibaba
TOKEN USECompare consumption on the same cases.

Use representative prior tasks so the candidate and current model are assessed against the same kind of work.

Compare both candidates on representative cases.

Telemetry & evaluation

Check the effect before changing a model.

Built-in Langfuse retains agent and system telemetry across tasks. On representative past cases, compare token use, latency and cost. Where reviewed outcomes exist, compare task quality too.

A faster or cheaper run is useful only if the result still meets the task’s requirements. Keep the comparison tied to the role and the cases that matter.

For builders

What to pin down in the configuration

Before swapping a model, inspect the agent’s job and the provider path that serves it. Exact identifiers and supported features belong in the release documentation.

01

Agent role

Which steps and decisions belong to this agent?

02

Provider & model

Which configured model and version is used for this role?

03

Access & tools

What context and tool calls does the role require?

04

Evaluation

Which past cases and reviewed outcomes will test a change?

Build with Corin

Try a team with different model choices.

Run a working example, inspect the agents and their handoffs, then assess a model change on relevant cases before using it for your work.