There is a false choice in this space: build every agent by hand on raw primitives, or hand your whole operation to a closed platform you cannot see into. We sit in the middle. We build the reliable backbone ourselves, on Cloudflare primitives, and we use open agent frameworks on top of it to move faster. Two come up again and again.
Hermes Agent, our default
The framework we reach for most is Hermes Agent from Nous Research. It is open source and self-hosted, which matters to us because it means no black box and no lock-in. What makes it stand out is the learning loop: it builds reusable skills from experience, keeps a persistent memory of past work, and carries context across sessions instead of starting cold every time. It is also model-agnostic, so we point it at whichever model fits the job rather than whatever a vendor allows.
In practice we use Hermes Agent when an assistant needs to get better with use and remember what it has done. Because it is self-hosted, it drops cleanly onto the infrastructure we already run, and we still front its model calls with a gateway so every call is logged and cost-controlled.
OpenClaw, for multi-channel assistants
The other one we use, less often but in the right spot, is OpenClaw. It is an open-source personal-assistant agent built around a gateway that bridges a long list of messaging channels, so one assistant can be reached over Slack, Telegram, WhatsApp, and more without wiring each one up by hand. When a client wants an assistant that lives where their team already talks, rather than in yet another dashboard, OpenClaw is a fast way to get there.
When we skip the framework
A framework is a head start, not a free pass. When an agent has to take real, irreversible actions in a client’s systems, we still build the core ourselves on Durable Objects and Workflows, because we want full control over state, retries, and what the agent is allowed to do. The rule of thumb is simple: use a framework to move fast on the parts that are well-trodden, and own the parts where a mistake costs the client something real.
What does not change either way is the discipline around it. Whatever runs the agent, every model call goes through a gateway for observability, fallback, and spend limits, and every agent ships with evaluation and monitoring. The tool is a means. A reliable AI agent is the point.