---
title: "Test Caveman alongside AgentOps in a multi-agent workflow"
description: "Start with one multi-agent task whose final result you can check. Keep AgentOps collection working while you test context reduction or a private Caveman Platform workflow. The migration should preserv"
canonical: https://caveman.so/switch/agentops
last-updated: 2026-09-07
---

# Test Caveman alongside AgentOps in a multi-agent workflow

Start with one multi-agent task whose final result you can check. Keep AgentOps collection working while you test context reduction or a private Caveman Platform workflow. The migration should preserve the connection between the task, its agents, and every attempt.

Caveman does not import AgentOps history through a public one-step tool. Arrange Platform access and agree on event support before replacing production collection.

## Save the instrumentation contract

Record AgentOps initialization, framework integration, session IDs, agent identities, operation decorators, and any manual event calls. Check initialization order against [the integration documentation](https://docs.agentops.ai/v1/integrations/agentssdk) for your SDK version.

Inventory dashboards, alerts, and debugging links that depend on those identifiers. Include scripts that calculate costs outside the UI. A field rename can break an investigation even when traces still arrive.

Preserve the working configuration and keep credentials available for rollback. Avoid adding overlapping automatic instrumentors until you have checked whether both wrap the same provider client.

## Build a task with a visible handoff

Use a staging workflow in which one agent collects evidence and another produces a final artifact. Include a tool failure and a retry. Keep external side effects disabled or confined to disposable fixtures.

Save the starting task, tool responses, framework version, model settings, and acceptance check. Record which agent owns each call and which attempt produced the accepted result.

Check totals at task level. A failed researcher attempt still belongs in the cost of the final article or patch. Do not drop it because the writer eventually succeeded.

## Prove recovery where agents actually run

A small fixture trial can establish whether Caveman produces a useful view:

```bash
npm install -g @caveman-ai/cli
caveman shrink --file ./fixtures/research-results.json
```

Use any emitted handle to inspect the original, following [the recovery steps](/guides/prompt-compression). Then test the receiving agent's access to that same content in the environment where the handoff runs.

If agents use different hosts, a local recovery handle is not sufficient by itself. Keep the original handoff content until you have an explicit supported recovery path. Do not copy a laptop's loopback URL into a remote worker configuration.

For a supported coding-agent trial on one machine, use `caveman setup --install`, run `caveman doctor` with the launcher name, and start that launcher through Caveman. This evaluates that session, not an entire remote agent fleet.

## Validate candidate collection

For a provisioned Platform pilot, agree on the source integration and mapping before sending records. Check these relationships on the same staging task:

| Relationship | Passing result |
| --- | --- |
| Task to session | One task's related work can be found together |
| Session to agent | Agent identity remains distinguishable |
| Agent to operation | Tool and model steps retain their owner |
| Retry to original attempt | Both remain visible without double counting |
| Usage to provider | Model and token semantics match the source |
| Final artifact to task | Reviewer can inspect what the agents produced |

Confirm what content is collected and where redaction occurs. If some records are sampled or absent, the resulting total is partial and should be labeled that way.

## Move one operational workflow

Use the candidate for one cost investigation or failure review before changing release checks and alerts. Have someone outside the implementation follow an expensive task from its total to the calls and artifact behind it.

Compare the candidate result against AgentOps and provider records. Resolve discrepancies in run boundaries, duplicate spans, missing prices, and retries before interpreting a lower total as a saving.

## Roll back

Restore the original instrumentor initialization and exporter settings. Keep the orchestrator, tool definitions, and agent identities unchanged throughout the trial so rollback does not alter execution.

Stop sending candidate records only after active sessions have finished. Preserve any recovery originals those sessions still need. Read [the comparison](/compare/agentops) to decide whether keeping AgentOps plus local Caveman is the right final setup.