Agents write the code now, and first drafts of the whole development lifecycle are next. Execution has never moved faster, but planning is another story. Product requirements crawl through weeks of coordination, review boards convene on quarterly calendars, and the artifacts that feed the agents are written by hand. That mismatch is the new bottleneck, and closing it means changing the lifecycle.
Palanivel Rajan Mylsamy, known as PVR, is Director of Engineering Program Management at Cisco and the author of The Agentic TPM, a field manual and agentic companion for engineering programs in the agent era. Behind it sits leading a multi-billion-dollar portfolio. Mylsamy's career spans over two decades across networking, cloud, and agentic AI. The book distills it into agent-era playbooks.
"The engineering lifecycle is shifting from the Software Development Lifecycle to the Agent Development Lifecycle. The intent is to give teams a field manual for building in the agent world without starting from ground zero," Mylsamy says.
The gap he targets sits upstream. Agentic coding is remaking the execution side, and the tooling trends keep compounding, but planning rituals barely moved. "Most of the agent tools we have are applied over execution to accelerate. Delivering things faster has never been the problem. My focus as a TPM leader is bringing agents into the planning lifecycle. The workflow remains the same, but we inject agents at each step."
His PRD generator shows the shape of it. "To write a PRD, you need customer research, UX thought through, a business case, and business value. You still need all of them, but it can happen with concurrent processing. Instead of having five to ten product managers start writing, I'm eliminating the operational overhead and getting the team reviewing days earlier." Meeting-prep agents do the same for commit reviews. "It takes three to four weeks to bring people together and get the content ready. As long as your content is organized in Jira or the Atlassian suite, my agents pull it ready, so you start the meeting at day 10 or day 20 rather than day zero." A product that took a quarter now turns around in two or three weeks.
The role changes too, echoing how every engineer is becoming a manager of agents. His proof is Atlas, a chief-of-staff agent from the book. "Every day I screen through five or six tools to get my baseline. My inbox, my calendar, my program dashboards in Jira, Slack, Webex. Easily it takes 90 minutes. Atlas runs a morning brief, scans all the tools, and sends a snippet to my Webex with a focused agenda, outcomes, and goals. I'm given my 90 minutes of work within a latency of 90 seconds, and I can take action in less than 10 minutes."
The reclaimed time compounds into orchestration at enterprise scale. "Earlier, if I could do two programs, I can run four, because I have concurrent processing and agents running for me. While I do four, I can add more value as a technical program manager."
Scale demands controls, especially as governance frameworks age out. "We consciously decide where to go all in on AI and where to keep the experts in the human-in-the-middle loop, so we still guardrail the decision when it comes to rolling out to customers," Mylsamy says. "Do we have the right security frameworks built in so we don't expose critical information? We don't want the LLMs to hallucinate the data. We need accuracy." Cost joins the list, feeding agent control planes and gateways that watch retrieval and spend. "A year ago we were all in the experiment phase, so we could spend whatever we wanted. We don't have that luxury anymore. That's where tokenomics comes in."
Two duties stay human. "Say you and I need to collaborate. Can the AI agents talk together and get us ready for this level of conversation? The answer is no. We are not there yet. It still needs contact eye-to-eye." Risk forecasting is the other. "I'm personally not comfortable yet, because LLM outcomes sometimes hallucinate. At least I'm not seeing it's there."
What separates his agents from automation is the architecture underneath, spanning deterministic control planes, data foundations for the agent era, and platforms that ground agents in enterprise data. Domain experts still pick the stack. "AI can provide guidance, but we as domain experts have to decide the right fit."
The book's six ADLC phases put context engineering at the front. "Intent comes first. The intents are key to building the right context. Then it generates, validates, orchestrates, and moves to operational learning," he says. "Using a generator to generate a document, anybody can do. It's just automation. There is no agent in it. Intent, context, generation, orchestration, and then deployment give you meaningful outcomes rather than a document without meaningful data."
