[2/2] Autonomy Is All You Need (Agents column)
Continue. story About the report by Michele Catasta, President & Head of AI at Replit, I would like to share the conclusions that can be useful to engineers from this report.
1“Autonomy” should be designed as a feature, not rely on a model If you are making your own agent/code assistant, it is important to take Michele’s position: Autonomy is not a property of a model, it is a property of a system.. We need to consciously build: Layer of automatic testing and validation Work models with repository and long context Planning/parallelization architecture
- The policy of kickbacks and errors (recovery) Otherwise, you get a very smart autocomplete, not an agent.
2️⃣ Autotests and CI/CD transform from engineering hygiene to agent API For development teams, this flips attitudes around tests and infrastructure: Good coverage with tests and fast CI is not only about people, but about how to do it. Agents could safely modify the system. Red → Green → Refactor becomes a cycle not only for the person, but also for the agent.
- Infrastructure. (test env, staging, feature flags) - That's it. Operating environment for an autonomous agentNot just a convenience for the developer.
If you want to trust an agent to do migrations, features, and refactorings in the future, you need to: Where to run the code in isolation How to make sure nothing is broken
- Where to roll back if it's broken
3️⃣ Context management as a new layer of product architecture Architecturally, “context management” for an agent is almost a separate service:
- Code and artifact index (vector + structural indexes); Long-term memory of decisions (design docs for the agent);
- Trajectory history. (What the agent did that worked that didn't); A planning layer that can: Cut tasks into subtasks
- Tracking progress Decide what can be done in parallel It’s a lot like adding an orchestrator to a microservices architecture, but now we’re orchestrating not services, but services. model.
4️⃣ Parallelism in agents = new UX and DevEx patterns For technical managers and platform teams: You need to think not only about how the agent “writes code correctly”, but also about how the agent “writes code correctly”. How the user experiences their work: Does the agent show understandable progress? Can the user intervene/adjust the plan? How parallel branches are displayed (logs, diagrams, "job view").
- Plan-oriented UI (Replit Agent, LangGraph-like systems) Developers want to see the trajectory of the agent, not the black box.
5️⃣ Strategic conclusion: AI-infrastructure will become the norm for girls If you take Michele's argument seriously, the next two 2–3 Years for engineers and techlids mean:
- We should invest in:
Testability/observability of the code;
Explicit domain modeling (So the agent has something to operate on.);
Infrastructure for experimenting with agents (sandbox, telemetry, safety‑rails).
Stop thinking of an agent as a “personal copilot”
agent teammatewhich: - Going backlog,
- Makes changes, The same quality gates pass as a person. (tests, reviews, linters).
#AI #ML #Agents #Software #Engineering #Architecture