Technical Pillars of the Agentic Dev Loop

Where do humans run code? Where do agents run code? When is execution local vs remote?

Should every task run in a fresh environment? How long do workspaces live? What state is allowed to persist?

How do we guarantee identical behavior across runs? How do we eliminate snowflake laptops?

Who is an agent, cryptographically? What credentials can it hold? How long do identities live?

Which agent handles which task? How is concurrency controlled? How are failures retried?

When are agents allowed to author code? How is authorship recorded? Who owns the result?

Can agents open PRs directly? How are branches named and scoped? How are tasks linked to specs?

Which reviews are automated? What do humans still review? How is drift prevented?

How do agents integrate with pipelines? When do they run? How are flaky behaviors detected?

What actions are forbidden? What violations are blocking? How is drift detected early?

Where must humans approve? How are confidence thresholds set? How are overrides handled?

What do we log? How do we trace actions across systems? How do we measure quality and cost?

Who did what? In which environment? With which credentials?

How do you write specs agents can execute against? When is a prompt enough vs. a structured charter?

How do you set up a local environment that agents can also use? What does the inner loop look like?

How do you swap your running dev environment with an agent? How does context transfer work?

How do agents access running services? How do you expose local or ephemeral environments to the outside?

How does an agent participate during the implementation phase? What does the human-agent handshake look like?

How do you use agents to build agents? What does the meta-loop look like?

Before any implementation begins, how do you verify the bug or behavior you are trying to fix actually exists?

After implementation, how do you verify the change works? What does the acceptance loop look like?

What evidence gets produced? Screenshots, logs, test results, diffs — how do you capture and store proof?

What does a real agent-authored PR review look like? Who reviews what, and when does a human step in?

How do you use the Claude Agent SDK to wire up agentic workflows? Patterns, gotchas, and practical examples.