Cloudflare's developer changelog describes a new way to run the OpenCode 2.0 coding agent on its platform: an experimental OpenCodeHarness in the Agents SDK that hosts OpenCode inside a plain Durable Object. The post, dated 12 October 2026, frames the harness as the mechanism that connects the agent runtime to Cloudflare's durable compute primitives rather than as a standalone product.
According to the changelog, sessions are stored in SQLite and Lifecycle jobs wake interrupted work after eviction. That pairing matters because Durable Objects can be evicted between requests; the documented design pairs persistent session state with a job that restarts work that was cut off, which is the part of the announcement with the clearest operational consequence for anyone running long agent tasks.
The post specifies the packages needed: Agents SDK 0.28.0 or later, installed together with the OpenCode packages via the command it prints, which pulls agents@latest, @opencode/sdk@2 and @opencode/plugin@2. Version floors and package names are the concrete, checkable part of the release; the changelog does not state a general-availability date or any pricing for the harness.
The documented setup imports DurableObject from cloudflare:workers, Lifecycle from agents/lifecycle, OpenCodeWorkerd from @opencode/sdk/workerd, OpenCodeHarness from agents/harness/opencode, and createAI from agents/models/opencode. A class extending DurableObject then constructs an AI binding, an OpenCodeHarness, and installs a Lifecycle that uses the harness. The changelog shows both a JavaScript and a TypeScript variant of the same example.
Inside the harness configuration, the example passes a storage object into OpenCodeWorkerd.create, sets models to fetch: false, and supplies plugins. The defaults block names a model, @cf/moonshotai/kimi-k2.7-code, with reasoningEffort set to low, and an agent named build. The class exposes an ask method that calls harness.prompt and returns status, text and reason.
According to the changelog, OpenCode reaches Workers AI via the AI binding by way of the agents/models/opencode provider. Custom tools, remote MCP servers and skills can be added to the harness through OpenCode plugins, the changelog adds. For the embedded host API, readers are directed by it to OpenCode's Cloudflare SDK documentation. A working example is also cited, one that brings in a browser UI, WebSocket streaming, and Workspace tools drawn from @cloudflare/computer for code execution and files.
For freelancers and small studios building agent-backed tools, the practical appeal is architectural: the agent loop, its session state and its model calls are described as living inside one Cloudflare primitive rather than in a separate always-on server. That is a supported reading of the changelog's own description, not a benchmark. The post supplies no latency, cost or reliability figures, so any efficiency claim would be unsupported.
The tradeoffs visible in the evidence are the ones the source itself flags. The harness is called experimental, which in Cloudflare's changelog vocabulary signals an interface that may change. The example pins @opencode/sdk@2 and @opencode/plugin@2 while installing agents@latest, a combination that can drift as the SDK moves. And the model default in the sample is a specific Workers AI model with a low reasoning effort setting, which readers would need to revisit for their own workloads.
There is also a scope limit worth stating plainly. The changelog documents how to wire the harness, not how well it performs. It does not report evaluations, comparisons against other hosting approaches, or the conditions under which Lifecycle recovery succeeds. The claim that Lifecycle jobs wake interrupted work after eviction is a description of intended behaviour in the source, not an independently verified outcome.
The announcement sits in a broader pattern of agent runtimes being packaged as SDK components rather than as hosted services, but the evidence here covers only this one changelog entry. Nothing in the supplied material establishes adoption numbers, third-party testing, or how the harness interacts with Cloudflare's other agent tooling beyond the imports and the linked example.
For this audience, the sensible next step is a contained experiment: install the stated versions, reproduce the Durable Object example, and check whether session persistence and post-eviction recovery behave as described under your own workload before depending on them. Treat the experimental label as a signal to keep the integration behind an interface you can swap, and budget for the possibility that package versions and the harness API change.
What remains unknown from the evidence: whether the harness is available on all Cloudflare plans, what it costs beyond normal Durable Objects and Workers AI usage, when it might leave experimental status, and whether the linked example is maintained alongside the SDK. The changelog answers how to start, not what happens at scale.