The Mainstream Picture

Much of what you read about agents describes something looser than the bounded loops in this course: a coding agent left to work in a repository for hours, an assistant that remembers everything about you and acts on your behalf, teams of agents handing tasks to one another, agents that improve themselves, companies where most of the work is agents talking to agents. It would be odd to finish this part without saying what to make of them.

The harness. Look closely at any free-ranging agent that works and you find a harness around it: a sandbox it cannot leave, permissions for what it may touch, a budget of time and tokens, a cheap way to check its work, and a person who reads the result. Coding agents are the success story of this pattern, and the reasons are specific. The work is verifiable at almost no cost, because the tests pass or they do not. Mistakes are reversible, because version control exists. The environment can be isolated. Where those three conditions hold, an agent can be given a long leash. Where they do not, the harness has to carry more of the judgement, and the leash shortens.

Many agents. Multi-agent systems split a task among specialised agents: one plans, several execute, one checks. Sometimes this genuinely helps, usually because each agent gets a smaller context and a clearer instruction. Often it is one agent's loop drawn with more boxes, adding cost, latency and a new failure mode, agents agreeing with each other's mistakes, without adding capability. The test is the same as for a single agent: what does each one read, what may it do, and who checks.

Self-improvement. What exists today is narrower than the phrase suggests: systems that run evaluations, notice failures and revise their prompts, tools or examples; models fine-tuned on their own filtered outputs; agents that write notes to themselves for next time. What does not exist in production is a model rewriting its own weights unsupervised, and the reason is the stopping-rule problem from chapter 12. A system that scores its own work needs something outside itself to say what good means.

The assistant and the agent-run company. A personal assistant is an agent with memory, identity and delegated authority, and its hard problems are exactly those three. The company run by interacting agents is the same picture at scale, and the questions it raises are the ones this part has asked: which agent did this, on whose authority, within what budget, checked by whom. An organisation that could answer those questions for a thousand interacting agents would already have built the structure this course describes. The mainstream picture and the structured one are not rivals. The picture is what becomes possible. The structure is what makes it safe to try.