[1/2] Autonomy Is All You Need (Agents column)
I saw an interesting one. report Michele Catasta, president & head of AI at Replit, told a month ago at the AI Engineer conference. Michele previously worked as the head of applied research at Google, and is now responsible for Replit’s AI strategy, which assembles application prototypes from scratch to demos in minutes. Here are the main points of his speech
1Autonomy is the main measurable progress in agents Code assistants can be evaluated not only by the quality of the clues, but by how far the agent goes himself, without a person on the hand. For non-technical users, this is generally the only meaning: either the agent is able to bring the task to the result, or the product is useless for them. Hence “north star”: degree of autonomy is a key metric for the development of AI agents, not just the quality of a single request.
2Two fundamental abilities for true autonomy Michele highlights two basic bricks of the autonomous agent in development:
1. Automatic testing The agent must be able to test himself through unit tests, integration checks, e2e scripts, health checks, etc. Without automatic validation, it either:
- Needs a permanent human reviewer.
Or it will “hallucinate” success and break sales.
In Replit, a whole cycle is built around this: code generation → test run → file analysis → auto repair. Without it, there is no real autonomy.
2. Advanced context management An agent who makes something more complicated than a single file must: Understand the structure of the repository and artifacts
- Keeping the state of long tasks (Days/weeks of project work) Remember decisions, compromises and limitations (memory) Manage the plan: what is done, what is broken, what subtasks are open
Without good context management, the agent either “forgets” important details through N steps, or begins to breed contradictions in the codebase.
3After Autonomy – Parallelism as the Key to UX When the agent can act on its own, the next challenge is how to make sure the user doesn’t have to wait forever. Michele analyzes several parallelization models: - Task‑level parallelism. Decomposition of work on independent subtasks: generation of front, back, configs, tests, etc. in different “branches” of execution. This reduces latency and gives early feedback: the user sees progress piece by piece rather than waiting for one giant response. - Out‑of‑order execution. It is not necessary to perform tasks strictly in the order of the plan, if there are independent pieces that can be dragged forward. A similar out-of-order idea in a CPU is to win over time, but you need to work carefully with dependencies. - Parallel plan-decomposition. Not one linear Chain of Thought, but a plan tree where different branches can develop separately and then collapse. This increases stability: you can roll back not “to the beginning”, but to the node of the tree.
Key idea: sequential agent = poor UX. The user gets stuck waiting and loses flow. A true AI engineer experience is when the agent rustles in parallel in several directions, and the person sees understandable progress.
4Balance: latency vs resources vs correctness * Once we add parallelism and autonomy, the classic engineering triple begins. Less latency → more parallel branches → higher consumption of tokens / calculations. More autonomy, less human control, more risk of incorrect change. Tough guarantees of correctness → more checks / manual confirmations → worse than UX.
Michele says there is no magic solution. You need to clearly design these three for your product:
- where we are willing to pay with computing resources for the wow effect;
- where for safety they agree to sacrifice speed;
- where you need an explicit point "here we always ask the person."
In continuation They're going to be thinking about what engineers can do to build their own autonomous agents.
P.S. By the way, the story of Replit was well told by Amjad Masad. (CEO) In an interview with Y Combinator in the summer (see mine analysis)
#AI #ML #Agents #Software #Engineering #Architecture