1 year ago working with AI agents often felt like a second job: babysitting, losing context, and cleaning up confident mistakes. The diagnosis was right.
The Anthropic research described agents working in shifts — read the git log, check a progress file, validate the current state, then code.
Split between an initializer agent and coding agents. That split helped. But the bigger unlock was making every handoff verifiable.
Back then, agentic coding still felt like a demo. Today it has become part of the default industry toolchain for many teams. The models did get better, but the bigger change was in the workflow around them. Planning PRD, RFC, use-cases, tests, code review, CI, merge requests, ticket updates — the agent now works inside the same guardrails we already trust for human engineers.
The pattern I keep seeing is simple: break the work into stages, and make each stage prove the previous one.
- A plan has to survive a review before implementation starts.
- A change has to survive tests and a code review before it reaches CI.
- CI has to pass before a merge request exists.
- Acceptance criteria have to be checked before the work is considered done.
The agent still forgets things. But it no longer has to hold the whole project in its head. The plan, the ticket, and the diff carry the context between stages — external memory with a schema, not a giant prompt.
The agent can still be wrong. But a wrong claim now has to survive several independent checks before it reaches main.
That's what changed. Hallucination didn't disappear — it got demoted from a silent failure to a caught failure. Babysitting didn't disappear either — it moved. Instead of watching every step, it is enought to compare the plan and the final change. Everything between them runs on evidence, not vibes.
The winning pattern isn't a smarter single agent with a giant context window. It is a system that treats the agent like a new engineer on shift: durable memory, a clear definition of done, and a chain of checks.
The teams that win don't have the smartest agent. They have the best verification harness.
Agents can replace the routine, not the responsibility.
We still need to bring our opinions, imperfect ones, our own ideas, objections and doubts, the thing you know that the model doesn't.
Top comments (1)
Babysitting didn't disappear — it moved to comparing the plan against the final change.