AI software factories still need humans: devs become gatekeepers of agent output
Pavan_Belagatti · x · 2026-11-16
Answering a recurring question — if agents automate the whole software pipeline, what's left for developers/SREs/DevOps? — the author argues these roles become human gates that verify and approve agent output at critical stages.
The accompanying AI Software Factory diagram lays out a six-stage agentic SDLC:
- Intake: agents triage tickets, Slack requests, alerts; owners auto-assigned.
- Plan & Spec: agents draft specs grounded in live catalog context; human gate on spec review.
- Build: coding agents ship features with AI skills, subagents, context files; scoped repo access.
- Review & Test: automated PR reviews and test runs scored against standards; human gate on PR approval.
- Release: release agents promote builds to prod only after human sign-off; hooks + CI/CD.
- Operate: event-driven incident response and vulnerability fixes, runbook-bound and audited.
Key point: faster coding ≠ faster delivery — the bottleneck shifts to review, testing, deployment, and ops, hence human gates at critical stages and runtime guardrails (least privilege, policy, token budgets, audit).
More from coding & agent
- Pragma Research's Cadence agent project demoed via a talking parrot — harrydong · 2026-11-16
- After 200+ client builds, a dev shares his Lovable design process as a free skill — damienghader · 2026-11-16
- Recreating a university Blender project in 20 minutes with Google Antigravity's Agents Swarm — gaganghotra_ · 2026-11-16
- Dev builds a game worth "a year of work" in one weekend with AI — majidmanzarpour · 2026-11-16
- Gary Marcus challenges dev: show symbols in your agent harness or you're confused — GaryMarcus · 2026-11-16
- 13 Claude Code 3D skill repos (310K+ stars) turn motion design into a writing skill — Roger_M_Taylor · 2026-11-16