Ollama Agent Mode Explained: What Typing ollama Does Now (v0.32)

Since Ollama v0.32.0 (July 11, 2026), typing the bare ollama command launches an interactive agent. Per the release notes, it lets you "chat with models, code, search the web, and delegate real work." The notes list no changes to ollama run or other model-running subcommands — every other invocation change in v0.32.0 is inside ollama launch (Codex→ChatGPT rename, simplified menu, deprecation warnings), covered below.

That's the short version. Ollama agent mode is the biggest change to the tool's front door since launch, and it landed two days after Ollama announced an $88M funding round. This page covers what the agent does per the official release notes, how your old workflows hold up, the Apple Silicon speedups that shipped alongside it, and the one privacy question worth asking before you hit enter.

---

What Typing ollama Does Now {#what-happens}

Update to v0.32.0 or later, type ollama with no arguments, and you get this (straight from the release notes):

❯ ollama
Ollama 0.32.0

▸ Chat, Code, & Work (glm-5.2:cloud)
    Chat with models, code, search the web, and delegate real work

The release notes describe it as "an agent to help you code and delegate work." In other words: the command that used to be a plain model runner now opens a session that behaves like a coding assistant — it can hold a conversation, write code, reach out to the web, and take on multi-step tasks.

Two details in that banner matter:

1. glm-5.2:cloud — the default shown is a :cloud model. The out-of-the-box agent points at cloud inference, not a model on your disk. More on that in the privacy section.
2. "search the web" — a live-web capability inside a tool most people installed specifically to keep things local. As of v0.32.1, using it prompts you to run ollama signin.

If you're new to Ollama entirely, start with our complete Ollama guide — this page assumes you already have it installed.

---

What Agent Mode Can Do {#what-agent-mode-does}

The release notes keep the description tight, so we will too. Per v0.32.0 and the v0.32.1 follow-up, the agent can:

The notes don't document flags, config options, or how to point the agent at a different model, so we're not going to invent any. When Ollama publishes those details, we'll update this page.

What we can tell you: agent-style work is brutal on weak models. Tool calling, multi-turn reasoning, and long contexts separate the models that can actually do this from the ones that just chat. If you want to run agent workloads against local models, see best Ollama models for agents — and note that v0.32.1 specifically "improved Gemma 4 tool calling and multi-turn reasoning," which tells you which model family Ollama itself is polishing for this. For wiring Ollama into external tools the standards-based way, there's our Ollama MCP integration guide.

---

Keeping Your Old Workflow {#old-workflow}

Here's the part most people searching "ollama new version agent" actually want to know: did my scripts break?

Based on the v0.32.0 release notes: your model-running workflows didn't. The notes list no changes to ollama run or other model-running subcommands — ollama run isn't mentioned at all, and no removal or breaking change to it is listed. The invocation changes in v0.32.0 sit in two places: the bare ollama command (now the agent) and ollama launch (the three changes below). So a script that calls ollama run is untouched; a script that called ollama launch codex needs updating.

Three adjacent changes worth knowing:

• Deprecation warnings on old models. ollama launch now warns before launching CodeLlama, Qwen2.5(-coder), Llama 3.x, Mistral, StarCoder, and base DeepSeek-R1 tags. It's a warning, not a block — but it's Ollama telling you those models are on the way out for agent use. If you're on one of them, our current model rankings will show you what to move to.
• Codex is now ChatGPT. The Codex App integration was renamed: it's ollama launch chatgpt now, with a --restore flag to return to your usual ChatGPT profile.
• Simplified launch menu. The ollama launch menu now only offers the most popular integrations.

If something does behave oddly after upgrading, the Ollama troubleshooting guide covers the usual failure modes.

---

The Web Search and Privacy Question {#privacy}

Ollama's entire pitch has been local-first. Agent mode complicates that in two documented ways, so let's be precise about what's confirmed and what's open.

Confirmed by the release notes:

• The agent's default banner shows glm-5.2:cloud — a cloud model, meaning default agent inference isn't happening on your machine.
• As of v0.32.1, "agent web search and fetch now tell users to run ollama signin" — search and fetch are tied to an Ollama account.
• The interactive agent "receives the current working directory."

Not documented (open questions): what gets logged when you search, how long it's retained, and how you point the agent at a purely local model. The release notes don't say, and we won't pretend they do.

The fair read: web search has to leave your machine — there's no local copy of the live web. The signin requirement is how Ollama gates it. None of this changes ollama run with local weights, which stays as local as it ever was. But if your reason for using Ollama is "nothing leaves this box," know that the agent's defaults, as shipped, don't match that posture — you'll want to check its configuration before using it for anything sensitive.

---

Apple Silicon: The MLX Speedups {#apple-silicon}

Agent mode is the headline, but the performance story underneath it is arguably worth the upgrade on its own if you're on a Mac.

• June 11, 2026: Ollama announced a new MLX engine for Apple Silicon — in its words, "Ollama's highest performance on Apple Silicon yet with MLX."
• June 30 (v0.31.1): Gemma 4 got multi-token prediction, making it "nearly 90% faster on average" on Apple Silicon. Ollama's June 29 blog post pitched it as "up to 90% faster" specifically with coding agents — which is not a coincidence given what shipped twelve days later.
• July 16 (v0.32.1): fixed MLX model cache leaks and made MLX text model loading respect OLLAMA_LOAD_TIMEOUT.

The pattern is deliberate: make local models fast enough on Macs that agent workloads — which burn far more tokens than one-shot chat — feel usable. A near-doubling of Gemma 4 throughput on Apple Silicon is exactly the kind of gain agents need.

Windows/Linux users didn't get nothing: v0.31.2 (July 6) enabled flash attention on older NVIDIA GPUs (compute capability 6.x) and improved iGPU offloading for vision models.

---

Version-by-Version Changelog {#changelog}

The last six weeks, compressed:

For the longer arc — every release back through the early versions — see our Ollama version history.

---

What the $88M Raise Signals {#funding}

On July 9, 2026 — two days before agent mode shipped — Ollama announced $88M in funding from Benchmark, Theory Ventures, 8VC, Y Combinator, and angel investors. The company says it serves 8.9 million developers. The announcement's title: "Ollama: all aboard open models."

Read the sequence and the strategy is legible:

1. Jun 11: MLX engine — make local inference fast on the hardware most developers carry.
2. Jun 30: 90% Gemma 4 speedup — make a specific model family agent-viable.
3. Jul 9: $88M raise announced.
4. Jul 11: the bare command becomes an agent, defaulting to a :cloud model, with signin-gated web search.

That's a model runner becoming an agent platform, with the account system (ollama signin) and :cloud tier as the obvious monetization surface. A free local runner doesn't need 8.9 million developers to sign in; a funded agent platform does. The blog title signals open models stay central — and nothing in the release notes walks back local inference — but the growth story investors funded is clearly the agent, not the runner.

---

Verdict {#verdict}

Agent mode is a bigger strategic shift than a practical one — for now. Practically: your subcommand-based workflows are untouched by anything in the v0.32.0 notes, Macs got a genuinely large speed bump, and the new default only bites if you type bare ollama expecting the old behavior.

Strategically, the ground moved. The default experience of the most popular local-model tool now points at a cloud model and includes account-gated web search. That's not a scandal — it's disclosed in the release notes — but it means "I use Ollama" and "everything runs locally" are no longer automatically the same sentence. Know which mode you're in.

Our recommendation: take the upgrade (the MLX and flash-attention gains are real), keep driving explicit subcommands for anything sensitive, and try the agent with eyes open about the :cloud default. If you're picking models for agent work, start with our agent-model rankings.

---

Sources {#sources}

• Ollama GitHub Releases — official release notes for v0.30 through v0.32.1 (all version claims above)
• Ollama Blog — MLX engine announcement (Jun 11), Gemma 4 speedup post (Jun 29), and the $88M funding announcement (Jul 9)

---

FAQ {#faq}