Show HN: Changed how I use agent harnesses

A Show HN post titled Changed how I use agent harnesses describes a practical shift in how one developer runs coding agents day to day. The author has used OpenCode for a long time and prefers it over paid agent products. Claude Code and Codex subscriptions do not make financial sense for them, so they stay on the free tier of OpenCode Zen for routine work and only reach for stronger models when planning or drafting requires more capability.

The friction is not model quality alone but provider mechanics under rate limits. Capable models come from Zenmux or similar providers, and those endpoints are heavily rate-limited. Mid-task, the workflow hits a ceiling: stop the run, open the config, swap the provider, restart the harness, then resume the same task. That sequence turns a continuous agent session into a manual handoff loop between config, restart, and recovery of context.

For engineers building with agent harnesses, the failure mode is operational rather than intellectual. The costly part is not writing the next prompt; it is losing flow when a free or cheap path is good enough for most of the work and a stronger, rate-limited path is needed only for planning and drafting. Anyone mixing free tiers with premium providers will recognize the same stop-swap-restart pattern and the incentive to redesign how models are assigned to phases of a task.

Market context is a stack of overlapping agent products with different pricing and limits. Claude Code and Codex sit behind subscriptions that this author rejects for cost. OpenCode Zen’s free tier covers everyday use. Zenmux-class providers supply the stronger models but throttle hard enough that a single long task can exhaust the allowance. Builders are already assembling multi-provider setups instead of committing to one paid agent product end to end.

The practical takeaway is to treat model choice as a phase-bound resource, not a single default for the whole session. Plan and draft on capable, rate-limited providers; execute on free-tier OpenCode Zen when that is enough; and reduce mid-task provider swaps so a limit does not force a full restart. What to watch next is whether more harness users adopt the same split-provider pattern, and whether tools make provider failover or phase-based model routing automatic instead of a manual config edit.