Ollama 0.32.0: the agent shell is now a tagged release

What changed

Ollama published v0.32.0 (11 Jul 2026). The headline is not another model catalogue bump — it is product packaging for the agent path that landed on main earlier this month.

What ships in the release notes:

• Interactive agent experience: running bare ollama now launches an agent shell for chat, code, web search and delegated work (demo default shown against glm-5.2:cloud).
• Integration rename: Codex App integration is now ChatGPT — ollama launch chatgpt, with --restore to return to a usual ChatGPT profile.
• Simpler launch menu: ollama launch keeps the popular integrations only; everything else remains reachable via the same command path.
• Deprecation warnings before launching older agent models: CodeLlama, Qwen2.5(-coder), Llama 3.x, Mistral, StarCoder, and base DeepSeek-R1 tags.

Separately on the default branch since the tag window, MLX model load timeout became configurable via the existing load-timeout env path (#14796) — useful if large Apple-silicon models were hitting the old hard 2-minute wait.

Release: https://github.com/ollama/ollama/releases/tag/v0.32.0

Why it matters

On 10 Jul we flagged Ollama’s agent TUI while it was still default-branch only and said: spike it after a tagged release. That release is now here.

Ollama is no longer only the local inference substrate. It is packaging the thin agent harness — model selection, tools, approvals, context handling — as the product you open when you type ollama. That collapses the “runner + separate agent framework” stack for lightweight local coding and research work.

It does not replace Hermes or OpenClaw. Those own scheduling, multi-agent routing, durable memory, integrations, and operational control. Overlap is real on the thin local path only: one model, a few tools, human-in-the-loop approvals.

The deprecation warnings also matter operationally. Teams still defaulting to Llama 3.x / CodeLlama agent profiles will get friction on launch. Plan model upgrades before you treat ollama as a daily shell.

My read

Worth a spike on a non-production machine.

Do not:

• swap Hermes/OpenClaw for bare ollama on anything that needs durable state, multi-agent routing, or ops controls;
• run --yolo-class auto-approve on machines with real credentials;
• assume the release notes mean every agent feature from main is polished — treat it as a first stable cut.

Do:

• install v0.32.0 side-by-side;
• run one representative local coding task through ollama agent, Hermes, and (if relevant) OpenClaw with the same model;
• compare tool-call reliability, approval friction, context behaviour, and recovery after failed shell/file ops;
• note the ChatGPT launch rename if you had Codex App automation scripts.

If Apple Silicon + MLX is in play, also verify large-model load times against the configurable timeout rather than assuming the old 2-minute hard stop.

Bottom line

v0.32.0 turns Ollama’s agent experiment into a product release. Spike it as a low-dependency local shell. Keep Hermes and OpenClaw for the work that actually needs an ops stack. Local inference vendors keep climbing the harness layer — plan for that overlap, do not pretend it is not happening.