Choosing an Agent model: start with the work you need

Want ClawPod Agent to summarize a document or prepare a draft, but unsure where to start with so many model names? Decide what work to assign, connect the AI service, and send one small request. Rather than finding the best model immediately, start by checking whether it handles your task well.

1. Choose the role the model will serve

The official Guide describes four model roles.[1]

Start with main, which handles the primary responses. Consider response quality, whether it can use the tools your work needs, and cost. A lighter model may suit sub, but do not assign sensitive judgments to it. These roles do not automatically determine the stages of document writing, so include your preferred sequence in the request.

Which model should heartbeat use?

Heartbeat is a feature that lets the Agent check its state periodically. Here, we are discussing an operating approach that goes beyond simple checks to continue work in progress and review results.

Because heartbeat resumes work in progress, reviews results, or makes recovery decisions, consider using the same model as main by default.

The recommendation above assumes an operating scope that includes continuing work and reviewing results. The current official Guide also recommends the same provider and model for main and heartbeat.[1] This is a recommendation, not a mandatory requirement. If you need a different configuration, review the required reasoning, tool use, and cost with your operator.

Whether checks run and how often depends on the Agent’s settings. Choosing a model or writing HEARTBEAT.md, the file containing check instructions, does not configure scheduled execution. Tell your operator what you want checked and ask them to confirm the execution conditions.[2]

2. Connect the selected model’s AI service

The AI service supplying a model is called a provider in settings. Credentials are the connection and authentication information that establishes access to that service. The connection method varies by service.

Check the current ClawPod Cloud Portal list for available models. Even if a model is listed, the Agent needs usable credentials. If the model you want is missing, ask your organization administrator rather than adding its name to an instruction file.[1]

If main and fallback use different AI services, both need to be connected. The official Guide advises checking each in a test Room. For a local model—a model on a server you operate—check an actual response from the Agent after testing the server connection as well.[1]

Before changing an existing Agent, record the current service and model for each role, along with credential names or identifiers only. Do not copy raw keys or tokens into conversations or notes. The Guide’s change procedure is to check the impact on work in progress, use “Save and restart,” and confirm that the Agent can run again.[1] Coordinate change authority and timing with your team first. If your screen or permissions differ, ask an administrator for help.

3. Check the response with a small document request

Paste the fictional notice below into a test Room instead of using real customer data. You can check whether the response follows the source without sensitive material or external actions.

Request to send — illustrative example

The following is a fictional notice for testing.

“The Springlight team holds a 30-minute weekly meeting at 2 p.m. on Thursday. Participants must add agenda items to the shared document by 5 p.m. on Wednesday. The meeting room has not been decided.”

Using only this notice, summarize ① the meeting schedule, ② what participants must do, and ③ what remains undecided, in one line each. Do not add missing information. Do not search externally, run tools, change settings, or send messages.

Expected result — illustrative example, not an observed execution

1. Meeting schedule: The weekly meeting starts at 2 p.m. on Thursday and lasts 30 minutes.
2. Participant action: Add agenda items to the shared document by 5 p.m. on Wednesday.
3. Undecided: The meeting room has not been decided.

Check that the response gives the meeting as Thursday at 2 p.m. for 30 minutes, the agenda deadline as Wednesday at 5 p.m., and the room as undecided. If it invents a room, mark that statement as incorrect and ask for a correction using only the source. Compare the corrected response against the same criteria.

This test checks response quality for this document. If your work needs tools, check those separately with a small, reversible action within your authorized scope. To identify the model actually used, refer to execution records your operator can inspect—not the model’s self-description in its answer.

Check fallback separately, too. A successful main response does not establish that fallback works. A separate response from the alternative model and an actual switch when main fails are different checks. If you need to verify switching, ask your operator for an approved non-production check. Do not damage production keys or deliberately cause an outage.

4. Ask an administrator if there is no response or a change causes trouble

If an error occurs, give your administrator the error text without secrets, the time, the target Agent, and the role you changed. This helps them check whether the selected AI service’s credentials are valid and usable by that Agent. For local models, also check server access and model readiness.[1]

If trouble starts after a change, do not change several roles at once. Use your pre-change record to decide with an administrator whether to restore the last configuration that worked. Check approval and work impact before restoring it, then check the Agent’s state and repeat the small request above.

When comparing models, use the same request to assess their answers and record cost and response time as well. The Guide advises tracking cost, errors, and latency in Analytics.[1] This one fictional document is enough for a first step. If the result is satisfactory, gradually extend your checks to small requests closer to the work you actually want to assign.

Sources

1. ClawPod Guide — Provider and model selection — Roles, AI service connections, change procedure, local models, and response checks.
2. ClawPod Guide — HEARTBEAT.md and periodic checks — Check instructions and execution conditions.

Source: ClawPod Guide — guides/select-models; guides/configure-heartbeat