🤖 Ollama Agent Mode (2026): What Typing 'ollama' Does Now
Ollama v0.32 turned the bare 'ollama' command into a local agent that codes, edits files and runs skills. What changed, how we use it, and how to opt out.
Derek Holt · Local AI & Hardware Writer
· 6 min read
⚡ TL;DR — quick answers
- What does typing 'ollama' do now?
- Since v0.32.0 (July 11, 2026) the bare command launches an interactive agent: multi-turn chat that can execute code in your working directory, read and edit files, and use web search when configured. 'ollama run <model>' still opens plain chat.
- Is Ollama agent mode safe to use?
- It executes code in the folder you launch it from, so start it inside a project directory, not your home folder. Since v0.32.4, skills the model tries to load itself require your approval; skills you invoke directly are trusted.
- Did agent mode change the Ollama API?
- No. The server still listens on localhost:11434 with the same OpenAI-compatible endpoints. Every script, editor integration and pipeline we had pointed at it kept working untouched.
For three years the bare ollama command just printed help text — the real work lived in ollama run. Since v0.32.0 (July 11, 2026), typing ollama in a terminal drops you into an interactive agent that chats across turns, executes code in your working directory, reads and edits files, and reaches for web search when you've configured it. We've been running it daily since the point releases stabilized, and this guide covers what actually changed, how the new Skills permission model works, and how to get the old behavior back when you don't want an agent touching your files. If you're brand new to Ollama, start with our complete guide — this one assumes you have it installed.
By the numbers
- v0.32.0 shipped July 11, 2026, replacing the bare-command help screen with the interactive agent (GitHub releases)
- Six point releases in under a month — v0.32.1 (Jul 16) through v0.32.6 (Aug 4) — iterated on agent directory context, download fixes and Apple-GPU speed (GitHub releases)
- v0.32.4 (Jul 25) added the Skills permission gate: model-initiated skill loads need your approval, user-invoked skills are trusted (remio.ai)
- 178.2K GitHub stars and 17.3K forks as of this writing (GitHub)
- The agent loop first appeared experimentally in v0.14-rc2 with a bash tool, an approval UI and a deny list for things like
sudo(Phoronix)
What actually changed — and what didn't
The change is surgical. Bare ollama = agent. Everything else is untouched: ollama run <model> still opens the plain REPL chat you know, ollama pull still pulls, and the API server still sits on localhost:11434 speaking OpenAI-compatible JSON. Not one of our pipelines — the batch-caption scripts, the editor integration, the Open WebUI front-end — needed a single edit after updating.
What you get in agent mode is a loop, not a chatbox: the model can propose a shell command or a file edit, you approve it, it runs, and the result feeds the next turn. The terminal UI also gained controls for inspecting and toggling the agent's system prompt, which is worth reading once — it tells you exactly what the model is being told about your machine.
A real session, start to finish
Our first serious test was a folder of ffmpeg batch scripts we use for AI video post-processing. We cd'd in, typed ollama, and asked why one script silently skipped files with spaces in the name. The agent read the script, found the unquoted variable, proposed the exact edit as a diff, and — after we approved — applied it and re-ran the batch to confirm. Three turns, no copy-pasting code in and out of a chat window. That loop is the entire point.
The official launch stream walks through the same workflow at length, and it's the best single video on what the agent is for:
Two habits we adopted fast: always launch from inside the project folder (the working directory is the agent's whole world, and v0.32.1 specifically improved how it uses directory context), and never launch it from your home directory — file-edit powers plus a 7B model's judgment is not a combination you want loose in ~.
Skills and the permission model
Skills are reusable capability packs — task instructions and code the agent can load on demand, following the emerging Agent Skills specification. The security question writes itself: skills can execute arbitrary code, so who decides when one loads? Since v0.32.4 the answer is a two-tier trust model. If you invoke a skill, it's trusted and runs. If the model decides mid-conversation to load one, Ollama stops and asks you first. That's the right default, and it mirrors where every serious agent framework landed in 2026: approval gates around file changes, command execution and external calls. Our rule on top of it: read a skill's code once before the first approval, then let it through.
Getting the old behavior back
If you just want yesterday's Ollama: use ollama run <model> — it is unchanged and opens the classic chat. Scripted and API workflows never see the agent at all, since it only exists in the interactive bare command. And if the agent concept isn't for you at all, LM Studio remains the GUI-first alternative that hasn't gone agentic.
Which local models make good agent brains
Agent mode is only as good as the model's tool-calling, and this is where hardware honesty matters. On our 16GB rig (the VRAM math lives here):
| VRAM | Agent brain we'd run | Reality check |
|---|---|---|
| 8GB | qwen3:8b | Fine for file Q&A, flaky on multi-step edits |
| 16GB (our rig) | qwen3:14b | The floor where the approve-run-fix loop feels reliable |
| 24GB+ | 30B-class tool-callers | Genuinely competitive with cloud coding agents on small repos |
Ollama itself now nudges you: v0.32.0 added deprecation warnings when older families — CodeLlama, Qwen2.5, Llama 3.x, Mistral, StarCoder, base DeepSeek-R1 — are used as agent brains. Heed them; we wasted an evening confirming the warnings are earned. If your card can't hold a 14B comfortably, a rented cloud GPU is the cheap way to test the next tier before buying hardware — a few dollars of RunPod time, as of August 2026, settles it. For pairing Ollama with a heavier external coding agent instead of the built-in one, this walkthrough is the setup we'd copy:
The point releases worth knowing
v0.32.1 (Jul 16) improved agent directory context and fixed an MLX cache leak; v0.32.3 (Jul 23) fixed stalling downloads; v0.32.4 (Jul 25) landed the skills permission gate; v0.32.5 (Jul 27) fixed a Metal bug on NVFP4 models; v0.32.6 (Aug 4) sped up Qwen3.5 on Apple GPUs and fixed OpenAI-format streaming. The Apple-silicon push got noticed well beyond the changelog:
Update straight to the latest point release — v0.32.0 on its own is the roughest version of this feature that will ever exist.
When to still use plain run
Honest limits: the agent needs web search configured before it can research anything, the skills ecosystem is weeks old, and no local 14B matches a frontier cloud agent on a large refactor. We still reach for plain ollama run for one-shot prompts, benchmarking, and anything scripted — and for maximum-control serving, raw llama.cpp still beats both. We also cut third-party agent stacks (n8n, LangGraph, AnythingLLM) from this guide deliberately; they wrap the API, not the new agent, and deserve their own piece.
How we tested: daily agent sessions on our RTX 4080 (16GB) Windows 11 rig across v0.32.3 through v0.32.6, with qwen3:14b as the primary brain, against real ffmpeg and Python utility repos — plus release-note verification against ollama.com and the GitHub changelog.
Our verdict
v0.32 is the biggest identity change in Ollama's history: it stopped being just a model runner and became a local agent environment, without breaking a single existing workflow. That last clause is why we're positive on it. The agent is genuinely useful at 14B and up, the permission model is more considered than most 1.0 agent features, and the off-ramp is one word: run.
Frequently asked questions
▸What does typing 'ollama' do now?
Since v0.32.0 (July 11, 2026) the bare command launches an interactive agent: multi-turn chat that can execute code in your working directory, read and edit files, and use web search when configured. 'ollama run <model>' still opens plain chat.
▸Is Ollama agent mode safe to use?
It executes code in the folder you launch it from, so start it inside a project directory, not your home folder. Since v0.32.4, skills the model tries to load itself require your approval; skills you invoke directly are trusted.
▸Did agent mode change the Ollama API?
No. The server still listens on localhost:11434 with the same OpenAI-compatible endpoints. Every script, editor integration and pipeline we had pointed at it kept working untouched.
▸Which model should power the agent?
A recent tool-calling model. On our 16GB rig qwen3:14b is the daily brain; older families (CodeLlama, Mistral, StarCoder, base DeepSeek-R1) now trigger deprecation warnings when used as agents.
The 5 best AI video finds, every week
New models, tested prompts, and what actually worked in our production — one short email a week. No spam, unsubscribe anytime.
Written by Derek Holt
Local AI & Hardware Writer
Runs the site's local-inference rig and benchmarks every GPU, quant, and speed-stack claim on it personally before it goes in a guide. Will not shut up about VRAM bandwidth.
Explore these topics
Every guide, comparison and prompt library we have on each.
Keep learning
Local LLMs · Ollama & LM Studio
ComparisonsLM Studio Bionic vs Ollama: Agent App or DIY Local Stack?
LM Studio's Bionic agent and Ollama solve different local AI problems. We compare agent features, GGUF and MLX engines, cloud spillover, and privacy controls.
