Chapter 2 · Part 1
Muse's brain: Muse Spark
Every step of the agent loop starts with the same question: given everything so far, what should I do next? Answering it is the model's job. Muse runs on Muse Spark, the first model from Meta Superintelligence Labs, released in April 2026. It also powers the Meta AI assistant.
Meta describes Muse Spark as natively multimodal. It takes in images as well as text, which matters more for an agent than for a chatbot. A photo of a receipt, a screenshot of a web page and a picture of a broken dishwasher are all input it can reason about.
Its most interesting feature for agents is that it can spend different amounts of thinking on a question. Scroll through the three modes.
Instant: the model writes its answer straight away. Fast and cheap, fine for easy questions.
Thinking before answering
A language model produces text one token at a time, and every token takes a fixed slice of computation. A model that writes the answer immediately gets only a little computation per answer. A reasoning model is trained to first write out working steps, often hidden from you, before committing to an answer. More tokens means more computation, and for math, planning or tricky judgment calls that usually means better answers. The industry calls this test-time compute.
The catch is that thinking costs time and money. Meta reports something neat here: they trained Muse Spark with a penalty on thinking length, and after a while the model began solving problems with far fewer reasoning tokens. Meta calls this thought compression. It's the same kind of skill as a student who stops writing out every line of arithmetic.
Thinking wide instead of long
The third mode, Contemplating, takes a different approach. Instead of one very long chain of thought, it runs several reasoning agents at once and combines their answers. Meta says this gives better results "with comparable latency". Four people working on a problem at the same time finish sooner than one person working four times as long.
You'll see the same idea again in Chapter 5, where Muse spreads real-world work across parallel subagents.
Why an agent cares
An agent runs the model many times per task, sometimes hundreds. If every step used the slowest, deepest mode, a simple errand would take ages and cost a fortune. So a well-built agent matches effort to the step:
- "Is this email a receipt?" gets an instant answer.
- "Which of these three phone plans actually saves money, given the fine print?" gets more thought.
- "Draft a negotiation strategy for this bill" gets the deepest thinking available.
Benchmarks compare models, but in a real agent the rest of the system matters just as much: which tools it has, what it remembers and how it's kept safe. Meta itself says Muse Spark still has gaps on long, multi-step agentic work. The rest of this course is about the parts around the model that make up for those gaps.
For how a model picks each next token in the first place, see How ChatGPT Actually Works.