What does the software development lifecycle look like when agents write most of the code?

Summary

• The stages stay: plan, build, test, review, release and operate. What moves is where people spend time: writing the spec, defining the checks, deciding what merges and setting the guardrails around agents.
• Spec-driven development means writing a spec before writing code with AI. The spec becomes the source of truth for the human and the AI.
• Coding suits agents because automated tests verify the output and give agents feedback to iterate on. Human review remains crucial for checking that a solution fits the broader system.
• Parallel agents produce many diffs. In Omnigent's Polly, one agent implements, a different agent reviews, each implementer opens its own pull request, and the human decides what merges.
• DORA's 2025 report calls AI an amplifier of an organization's existing strengths and weaknesses. Unity Gateway adds the control layer: access, rate limits, budgets, usage records and traces for coding agents.

What does the software development lifecycle look like when agents write most of the code?

The stages stay the same: plan, build, test, review, release and operate. What changes is where people spend their time. Agents take over more of the typing, so people concentrate on intent (the spec), the definition of correct (the tests), the merge decision and the guardrails around what agents can reach and spend. DORA's 2025 report describes AI's primary role as an amplifier of an organization's existing strengths and weaknesses, and says the greatest returns come from the underlying organizational system, not from the tools themselves.

How does each stage change?

Stage	What changes	What stays with people
Plan	A written spec comes before code, and context files such as AGENTS.md or CLAUDE.md tell agents your conventions	Intent and acceptance criteria
Build	Several agents work in parallel, each in its own git worktree	The goal and when it is done
Test	Automated tests give agents feedback to iterate on	Defining correct behavior
Review	A different agent reviews the diff, and blocking issues loop back as fixes	The merge decision
Release	Agents open pull requests, and CI jobs can run code review and issue triage	Approval
Operate	Usage, spend and agent activity are recorded per user	Budgets, access and policy

Why do specs and tests carry more weight?

Birgitta Böckeler of Thoughtworks defines spec-driven development as writing a spec before writing code with AI, where the spec becomes the source of truth for the human and the AI. Her experiments also show the limits. She frequently saw the agent not follow all the instructions, and generating code several times from the same spec gave varying results. That is where tests come in. Anthropic lists several reasons coding suits agents, among them that code solutions are verifiable through automated tests, agents can iterate on solutions using test results as feedback, and output quality can be measured objectively. Anthropic adds that human review remains crucial for ensuring a solution fits the broader system requirements.

What does review look like when an agent wrote the diff?

Omnigent's Polly shows one workable pattern. It breaks a task into sub-tasks, delegates each to a sub-agent on its own harness and git worktree, and sends each diff to a reviewer from a different vendor. Each implementer opens its own pull request, and Polly never merges. Omnigent is open source with status alpha, and in Beta on Databricks.
Around that loop, Claude Code hooks can run a shell command before or after an action, such as running lint before a commit, and Claude Code can automate code review and issue triage in CI with GitHub Actions or GitLab CI/CD.

What controls does the lifecycle need around the agents?

Unity Gateway governs external coding agents such as Claude Code and Codex when you route them through a Databricks model service.

• Access: Unity Catalog decides who can call which models, through EXECUTE on a model service.
• Volume and spend: rate limits cap requests or tokens per minute, and monthly budgets can alert or block at per-user thresholds.
• Records: system.ai_gateway.usage records requests by requester and model. The usage dashboard's Coding Agents tab shows active days, coding sessions, commits, and lines of code added or removed.
• Traces: unified tracing (Beta) adds client-side traces from Claude Code and Codex to the gateway traces in one Unity Catalog table.
• Instructions: Genie Code, the Databricks coding agent, reads AGENTS.md and CLAUDE.md files in your workspace.

FAQs

Do developers still review code?

Yes. Anthropic's guidance says automated testing helps verify functionality, while human review remains crucial for ensuring solutions align with broader system requirements. In Polly, the human decides what merges.

How are specs different from AGENTS.md?

Böckeler separates specs, which apply to the tasks that create or change one piece of functionality, from general context documents such as rules files, which apply across all AI coding sessions in a codebase. AGENTS.md is a README for agents, so it is the second kind.

Does this only apply to Databricks tools?

No. The pattern fits any coding agent. Databricks' AI governance guide says Unity Gateway governs external coding agents, such as Claude Code and Codex, when you route them through a Databricks model service.