Agentic SDLC: Human-Led AI Throughput
Executive Summary
Most software teams still build the way they did before AI agents existed. Work moves from a PRD to epics and stories, through sprints, QA, and release, and every step is done by hand. Each hand-off depends on the person who wrote the last artifact, and knowledge of how the process really works lives in people's heads.
Alethea, a technology organization, came to Tech Holding to change that for its product and engineering team. Over an 8-week consulting and training engagement, one embedded Tech Holding consultant worked alongside Alethea's product managers, engineers, and QA staff to redesign their existing development process as an Agentic SDLC. Each stage is now supported by purpose-built Claude agents and skills, and each person acts as a conductor who directs and approves agent work rather than producing every artifact manually.
The engagement was not a platform build, and it was not about replacing people. It left Alethea with a process map tailored to how the team works, a version-controlled library of agents and skills, Jira and GitHub integration, role-specific training, and a team that proved it could run the workflow without the consultant, on new builds and maintenance work alike.
The Problem: A Manual Process That Could Not Scale
Alethea's process worked. Teams received a high-level initiative, internally called a "big rock," aligned on design and a PRD up front, broke the work into epics and user stories, ran two-week sprints, and carried it through QA, deployment, and post-launch monitoring. The problem was not that the process was broken. It was that every stage relied on manual effort:
- Manual, time-intensive stages: every step from PRD through deployment was done by hand, which slowed velocity and made the quality of hand-offs between phases inconsistent.
- No shared agent or skill library: without a common repository of agents and prompt skills, anyone experimenting with AI started from scratch, and the organization built no lasting AI leverage.
- No role-specific training: product managers and engineers had not been trained on agentic workflows suited to their part of the lifecycle, and generic enablement rarely changes how people work day to day.
- Undocumented process knowledge: how the team actually worked was informal and unwritten, so it could not scale and new hires could not onboard quickly.
- Risk of a one-off effort: without a plan to apply new AI workflows to ongoing features and bug fixes, any change risked being dropped after the first project instead of becoming the standard way of working.
The result was a capable team whose output was capped by how much manual work each person could get through in a sprint.
Solution: An Agentic SDLC Built Around the Team
Tech Holding embedded one Agentic SDLC consultant with Alethea's team, first on-site in Midtown New York and then remotely. Rather than imposing a fixed template, the consultant started from how Alethea already worked and built the agentic version of that process together with the people who would run it.
- Process Mapped – Discovery sessions with PMs, engineers, QA, and leadership mapped the current workflow, pain points, and hand-offs. The team then adapted the Agentic SDLC reference framework to its own scale and project types.
- Roles and Gates Defined – Each phase was assigned an owner (PM, engineer, QA, or tech lead), with a human approver at every gate so people stay in control of what moves forward.
- Agents and Skills Built – The consultant and team built phase-specific Lead Agents, supporting agents, critique agents, and reusable prompt skills, using off-the-shelf components where they fit and custom ones where the workflow required them.
- Live Project Run – A real project ran through every phase, from opportunity intake to post-launch analysis. Each role owner executed their own phases while the consultant guided and corrected in the actual workflow, not a classroom.
- Maintenance Proven – A feature addition and a bug fix went through the same workflow, confirming the process applies to ongoing work and not just greenfield builds.
- Independence Validated – Each trained team member ran their phase workflow alone in a simulated project. A month of post-handoff support followed, with bi-weekly check-ins and minor refinements.
Product Management (Phases 0–2)
Product managers run opportunity intake, market discovery, and PRD generation through an interview-driven agent flow. A PRD Interview Agent draws out requirements, and a Senior PM Critique Agent reviews drafts before hand-off, so the PRD that reaches engineering is complete and consistent.
Engineering (Phases 3–6)
Engineers use an Architecture Lead Agent for solution architecture and technical design, an Epic Decomposition Agent and Estimation Skill for delivery planning, and Claude Code for AI-assisted implementation backed by a Code Review Agent. Reusable skills such as the User Story Skill and API Contract Skill keep artifacts consistent from project to project.
QA & Release (Phases 7–9)
QA and release owners generate test plans, run regression agents, plan deployments, and analyze post-launch telemetry with agent support. Every phase produces a well-formed artifact that becomes a clean input to the next.
Jira & GitHub Integration
Delivery planning outputs flow directly into Jira as stories, and implementation agents reference the GitHub codebase for context. The full agent and skill library lives in a version-controlled repository that Alethea's team owns.
Results at a Glance
8 weeks from kickoff to a self-sufficient team | 10 SDLC phases supported by agents | 3 roles trained on their own phases | 0 roles replaced by automation |
The number that matters most is the last one. The goal was never to automate people out of the process. It was to change what their time goes toward. PMs, engineers, and QA staff still own every phase and approve every gate, but agents now produce the first draft of each artifact, from the PRD to the test plan. Because training was role-specific and confirmed through a proof-of-independence exercise, the know-how sits with Alethea's team rather than with the consultant.
Benefits
- Throughput Without Headcount Changes: agents handle first drafts and routine work at every stage, so the same team can take on more without adding or removing people.
- People Stay in Control: every phase has a human owner and approver. Agents propose; people decide what moves forward.
- Consistent Hand-offs: each phase produces a standard artifact that feeds the next, cutting the rework that comes from incomplete PRDs, designs, or test plans.
- Role-Specific Enablement: PMs, engineers, and QA staff were each trained only on the phases they own, backed by separate guides and short walkthrough videos for every phase.
- Documented, Repeatable Process: a complete Agentic SDLC Process Guide, artifact templates, and a version-controlled agent and skill library make the process easy to maintain and fast to teach new hires.
- Built for Ongoing Work: proving the workflow on a feature addition and a bug fix made the Agentic SDLC the team's standard operating model, not a one-off experiment.
Tools & Technologies Used
- Claude Code: AI-assisted implementation, code review agents, and command-line execution of agents and skills inside the development workflow.
- Claude Desktop: interview-driven PRD generation, discovery, and planning work for product managers and tech leads.
- Jira: destination for delivery planning outputs, with epics and stories generated directly from the planning phase.
- GitHub: codebase context for implementation agents and home of the version-controlled agent and skill repository.
Our Partners
By using this site, you agree to thePrivacy Policy.