Jennifer Artabane is SVP of Product Management at Auros, with 20+ years of experience in enterprise software, health tech and e-commerce.
Every serious software company supports Model Context Protocol (MCP) now. In under two years, MCP has gone from experimental to an assumed offering. According to a December 2025 Stacklok survey, almost half of the responding technical leaders from the software industry stated they were using MCP in production.
MCP support on its own no longer distinguishes anyone. Best-in-class MCP is now an expectation, and the question is now what your agents can reach through MCP? For product teams, the answer that matters is real customers and the ability to learn from them.
While agents have made building fast, they haven’t accelerated learning from customers in the same way. The teams that pull ahead are the ones that fold customer insight into every stage of the AI product development life cycle instead of running it as a separate step tacked on at the end.
The Missing Half Of The AI-PDLC
For 20 years, product development has been getting quicker. The AI-native product development life cycle (AI-PDLC) is the latest turn of that screw. Agents draft specs, generate prototypes, write code and run QA. The stages are the same, but the speeds at which they’re completed aren’t, with AI and humans now working together on those tasks. Work that used to require a sprint can now fill an afternoon.
The same is happening to research, validation and feedback loops, going from cycles measured in weeks to cycles that can happen in hours or days. Human feedback needs to be rigorous but doesn’t need to slow down development. The goal is to learn at the same pace you build so you can keep moving fast without losing sight of what you’re building and why you’re building it in the first place.
If a build only takes hours, a feedback study with customers at the end of the life cycle that takes too long will offset any speed gains your AI-PDLC earned you. However, that study is nonnegotiable. It has to be done. This juxtaposition wasn’t present in previous development shifts because AI output looks finished but is rarely complete and must be validated.
When Research Becomes Something Agents Call On
We worked with Zigzag, a puppy-coaching app, to develop and refine its AI chat feature, Ziggy. The team ran numerous tests throughout development using real-time customer feedback to decide which features were ready for release. In the first six months, Ziggy had received a 96% positive customer rating, been used over 80,000 times and helped cut human trainer support requests by 27% thanks to constant, high-quality user feedback.
Customer feedback is most valuable when it comes back to a product team in a time frame that enables it to iterate rather than rebuild the product. This is even more true if agents and AI are involved in what you’re developing or making; that end-user feedback will make or break it, and it needs to come in quickly.
However, gathering that feedback is now shifting from a stage in the life cycle to a capability available at any point in it. MCP has made this possible. An agent can put a question to real people mid-task the same way it queries a database or pulls a document. Research-based MCP servers let teams recruit participants, create studies and launch tests from Claude, ChatGPT and all other LLM-based tools, such as Figma Make, without leaving that tool.
When results are ready, they can be accessed from the same place, making it easy to incorporate into next steps. This approach changes how the research is completed and actioned but also who it’s for; increasingly, the primary users of customer research are people and agents in equal measure.
Humans Staying In The Loop
The obvious shortcut, tempting as it may be, is to simulate the feedback with synthetic users only. Synthetic user insights can be useful, but companies should exercise caution. Models are built to produce the most likely answer, so asking one model to validate another’s output may confirm what’s typical and miss what’s surprising.
The moments of surprise are where the product value sits, so don’t forgo the user or customer.
Start by mapping where in your product development you want to reach a real customer. Most teams find agents own the build steps entirely but have no route to customer input until a human commissions a study. This is where the change needs to happen, giving agents access to that human insight early on.
Still, remember to treat agent access to people as a governance decision rather than a feature switch. For that, you’ll need rules for spend (so recruitment costs don’t run unchecked) and rules for approval (so a person signs off before an agent launches anything).
Good governance is essential, and I’d suggest product leadership keep three questions in mind here:
1. At what stage of our life cycle can an agent ask a real customer a question?
2. Who decides when that’s allowed?
3. Are we measuring how fast we build or how fast we learn?
How To Learn As Fast As You Build
MCP support became table stakes within two years. Agent-driven build speed is heading the same way. Once every team can build in hours, the advantage belongs to the organization that incorporates real customer insight into its AI-PDLC and gets it back quickly. Agents are helping to shorten the distance between making a decision and getting evidence from the people it affects.
Identify what decisions you want agents to make, give them the right guardrails, connect research results to the build and measure how quickly you made it happen—all while keeping people involved at every level. The teams that build research into every stage of the AI-PDLC and keep real people in the loop as agents take on more of the work will set the standard for how products get made next.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?