Agent Skills are simple documents that tell a model where to look. In practice, a skill is a folder of reference material and scripts that an agent loads on its own once it recognizes the task in front of it. Boomi launched Companion in May, and the suite now covers five of them. It doubles as a working example of what an Agent Skill looks like in production, and a playbook for any enterprise vendor chasing the same default status inside a developer's AI agent, before a competitor gets there first.

What Boomi Companion Is

Companion is five open-source Agent Skills. Agent Skills are an open standard, published at agentskills.io, for packaging knowledge that an AI agent loads on demand (Boomi Developer Documentation, 2026). Each skill is a folder of platform specifications, connector schemas, error patterns, and best practices.

The five skills cover five jobs: building integration processes, searching the Boomi Marketplace for recipes, managing Data Hub models, deploying Agentstudio agents, and building Boomi Data Integration flows.

The standard is open. Companion runs inside Claude Code, Claude Cowork, OpenAI Codex, GitHub Copilot, and more than thirty other compatible agents (Boomi Developer Documentation, 2026). Claude users get a packaged plugin with slash commands and sub-agents. Everyone else installs the standalone skill.

The skills also include platform tools, scripts that push and pull components, deploy processes to a runtime, and run tests, using credentials stored in the developer's workspace. Companion works against a live Boomi account.

Why This Matters Beyond Boomi

Agent Skills do not turn themselves on. A developer has to install one first. So publishing a skill does not block a competitor from an agent's attention. It races to be the skill already sitting in the workspace before the developer looks for another one.

Once a skill is installed, it stays loaded. Every future request on that project gets built against the vendor's own specs, not whatever the model half-remembers from training data. That is the win, default status inside the agent, earned once, at install, kept through every session after.

The advantage grows for vendors who sit at more than one layer of the stack. Boomi's five skills span three layers, integration, data management, and agent orchestration. Red Hat touches compute and network. Cisco touches network and security. Either could ship one skill per layer and claim the same ground.

How to Implement This

1. Map your layers before you build. One layer, one strong skill. Multiple layers, one skill per layer, not a bundle. A developer working on one job should not have to load context for another.
2. Convert what you already have. Component specs, connector schemas, and known error patterns already sit in your developer portal. That is the raw material.
3. Publish on the open standard. A skill locked to one AI agent solves a smaller problem than one that follows the developer to whatever tool they use.
4. Decide read versus write before you ship. A skill that explains how to build a connector is low risk. A skill that deploys code straight to a live production environment is not. That call belongs to security and product, not DevRel.
5. Keep each skill narrow. Overlapping descriptions inside one bundle make the agent guess which skill applies.
6. Version against your releases. A skill needs the same release discipline as the product it describes. Name an owner to catch staleness before a developer hits it.

Risks to Manage

Write access is the biggest one. An agent that explains how a connector works is safe. An agent that deploys untested code to a live runtime, using stored credentials, raises governance questions a documentation skill never faces. Vendors with skills across multiple layers carry more of this risk.

Overlapping skill descriptions are the quieter one. If two skills both claim to handle a task, the agent has to guess which applies, which is exactly why step five matters before launch, not after.