local-agent-bench results

Which local Ollama models can drive a tool-using agent? This is every run from a small benchmark that tries to answer that. Five local models ran the same 15 agent tasks, 3 times each, with and without a one-sentence hint in the tool descriptions. That makes 450 runs.

The agent has three tools. It can search the web, read a page and use a calculator. It's one LangGraph agent loop over an MCP server, running on Ollama.

• Code: https://github.com/iamsorenl/local-agent-bench
• Write-up: https://larsensoren.com/writing/local-agent-bench/

Results

Model	Hints	Passed	Research	Median s/task
gpt-oss:20b	no	43/45	18/18	6.1
gpt-oss:20b	yes	40/45	16/18	5.9
gemma4:12b	no	33/45	9/18	27.9
gemma4:12b	yes	33/45	9/18	27.2
qwen3-coder:30b	no	27/45	6/18	3.0
qwen3-coder:30b	yes	33/45	9/18	2.2
gemma4:26b-a4b-it-q4_K_M	no	27/45	6/18	17.3
gemma4:26b-a4b-it-q4_K_M	yes	30/45	6/18	18.1
llama3.1:8b	no	24/45	3/18	3.1
llama3.1:8b	yes	24/45	3/18	3.3

The main finding is that most models fail at the same step. The search snippets in the test never contain the answer, so the model has to open a result with read_page. gpt-oss 20B did that in all 18 research runs. The others did it in 6 to 12 of 18, and the rest of the time they searched again and gave up. No model followed the planted prompt injections in any run.

Files

results.jsonl has one row per run.

Field	Meaning
model	Ollama model tag
variant	baseline, or hints for the run with one extra sentence in two tool descriptions
task, cat	Task id and category (math, research, chat, no-result, injection)
rep	Repeat number, 0 to 2
pass	True when every check passed
checks	The individual checks. finished, answer, called (required tools used), not_called (forbidden tools avoided), no_injection
calls	Tool names in the order they were called
blocked	Exact repeat calls the agent's guard refused
text_call	The model wrote a tool call into its reply instead of making it
out_tokens	Output tokens across the run
seconds	Wall-clock time for the task
answer	The model's final answer
error	Crash or timeout message, otherwise null

tasks.jsonl has the 15 tasks, with the question, the accepted answers, required and forbidden tools, and the planted injection markers.

How it was run

Every place and fact in the tasks is made up, so no model can answer from memory. Search results and pages come from recorded fixtures instead of the live web, so every model saw exactly the same content. Every model used the same agent, at temperature 0 with 32K context, a 2048-token output cap, a 20-step limit and a 5-minute timeout. Everything ran on one Apple M3 Pro with 36GB and Ollama 0.34.2, in September 2026.

The pass values are re-scored with the current answer checker. A curly-apostrophe bug failed a few correct answers in the original run, and fixing it moved gpt-oss 20B baseline from 41 to 43. Re-scoring can only turn a fail into a pass.

Limits

15 tasks is enough to show the pattern but not to rank models that are close together. Only 2 tasks test injection. At temperature 0 the three repeats agreed in 147 of 150 cases, so this is closer to one run per task. Answers are checked by phrase matching, which can miss a correct answer worded differently.