By Ricky Kirkendall
Choosing an agentic harness is one of the most consequential decisions when building and maintaining productive agent teams. Popular harnesses such as Anthropic's Claude Code and OpenAI's Codex are tightly coupled with their providers' models, so choosing a harness means evaluating not just its technical capabilities, but also its provider as a long-term partner. Key considerations include:
These are reasonable expectations of an infrastructure partner, even in a market that's moving this quickly. But when your harness ties you to one provider, each product announcement can leave you reassessing the economics or long-term viability of your setup. That uncertainty can become a distraction from building automations that deliver lasting value.
Our approach at Crafting is to build the harness into our platform so that our customers retain optionality on models and provider partners as the landscape evolves. Our success depends on helping customers build lasting value, with the freedom to continually evaluate and refine their automation stack based on solution merit rather than technical lock-in.
As your team works with an agent, they learn when to trust it. They discover what it can and can't do, how much detail a request needs, and how different models affect the results. Over time, that experience becomes a practical understanding of how to work with the agent effectively.
A harness and model from the same provider can work especially well together out of the box, but adopting that stack also ties your team more closely to that vendor. Moving agent context files between harnesses may be straightforward; rebuilding confidence in how a new stack handles your team's work takes time and experimentation.
We believe that investment in learning and trust-building should continue to pay off as your agent setup evolves. Changing models still requires evaluation, but it shouldn't also mean replacing the harness and relearning how to work with it.
We built our harness into the sandbox platform we've spent years developing for engineering teams. Agents inherit that foundation, including the environments, connectivity, and access controls needed to work across real systems.
A Crafting sandbox contains the workspaces where agents run, but it also gives them a way to work with the rest of your infrastructure. Agents can create new sandboxes and work within them. Their environments can use Kubernetes interception, sandbox access controls, and identity federation, with access governed by the permissions you've set up. The native harness has the platform's tools available to it from the start.
That's a big part of why we built our own harness. Existing harnesses work well within a workspace, but Crafting sandboxes extend beyond that. We wanted agents to make full use of the infrastructure and capabilities we'd already built.
Our Mac sandbox project illustrates this: the agent had the repos, CLI, and AWS access it needed to work through the implementation and test against real infrastructure. Getting those things into the environment makes a huge difference in how much work you can actually hand off. That's why positioning the harness at the sandbox layer makes sense to us: customers can change models and evolve their agent workflows without rebuilding the environment that makes it all possible.
We support OpenRouter natively in Crafting's LLM configuration. It's a familiar starting point for people who want to try open-source models, and you can use it with Crafting agents without needing a different system for the experiment.
Self-hosted models work too. You can connect your own inference endpoint to Crafting using a supported API format, such as an OpenAI-compatible API. That includes compatible Ollama setups, giving you another way to use open-source models alongside closed-source options.
We also support AWS Bedrock and Google Cloud Vertex AI, so you can keep using the providers you already buy inference from. The provider setup guide covers configuration and authentication. Admins manage providers centrally, while model aliases and purposes let them change model selections without requiring each developer to update their setup.
Building effective agents takes iteration. Our agent evaluation example shows how agents in Crafting can evaluate other agents on the same task, helping you compare prompts and models as you refine your setup.
As companies move from working with individual coding agents to running teams of agents that coordinate more autonomously, getting the foundation right becomes even more important. There's more to configure, more access to manage, and more to understand about how the system behaves.
Agent teams can automate larger parts of a workflow, but trusting them takes more than confidence in each agent on its own. Your team also needs to understand how agents divide the work, build on each other's results, and respond when something goes wrong.
Greater autonomy should build on what your team has learned, not force it to start over. We built Crafting to give companies a stable foundation for that progress: the tools and execution environments to earn trust over time, with the freedom to choose the models that work best. That combination gives teams a path to greater autonomy: adopt better models as they emerge, carry forward the workflows you've proven, and steadily expand the scope of work you can confidently hand off.
Happy Crafting.
