Phase 2ADLC course

Phase 2. Discovery: the agent that talks to stakeholders itself

We have a value hypothesis from phase 1. Now it has to be checked against reality — go to the people who supposedly feel the pain, and to the stakeholders the product depends on, and find out what's actually needed. This is discovery: interviews, custdev, context gathering.

This is where the course, for the first time, sends the agent out to live people. Until now it worked with data and artifacts. Now it initiates contact itself: writes to strangers, schedules meetings with them, talks, transcribes, and synthesizes. This is a role conventionally treated as empathic and therefore immune to automation — "people should be listened to by a person."

The human role today

A business analyst, product owner, part of the PM role — whoever finds out what's actually needed. They find people who feel the pain and stakeholders who matter, arrange interviews, run them, listen, transcribe, and distill what they heard into structure: segments, pains, the jobs people "hire" the product to do (JTBD), contradictions between stakeholders.

The bottleneck here is human throughput. Interviews are long, there are physically few of them, the interviewer gets tired, forgets half of it, and the summary is subjective — shaped by what they remembered and what they already believed. Discovery almost always runs into the quantity of conversations and the losses on the way from conversation to insight, not the quality of any single one.

What we hand to the agent

The agent runs discovery as a whole role. It decides who's worth asking under the opportunity-brief hypothesis, formulates questions, initiates contact, conducts or transcribes the conversation, distills what it heard into structure, and — critically — loses nothing along the way. Every word of every interview stays in the system, not in the memory of a tired interviewer.

The economics of the phase itself change. Where a human runs five interviews a week, the agent runs fifty in parallel, without degrading by end of day and without the bias of "hearing what I already believe." Discovery stops being a bottleneck.

Agent architecture

Phase state-machine

Inputs

opportunity-brief (phase 1), a map of stakeholder types and pain-bearer types, access to communication channels and calendar.

The agent holds the role

Tools: meeting scheduler (books calls itself); email and messaging (sends questions, arranges meetings); transcription; insight synthesis; deduplication and pain clustering; a contradiction detector across sources.

Artifact

discovery-report — segments, pains, JTBD, contradictions between stakeholders, a list of open questions, and a confidence level for each finding.

Handoff: discovery-report → Marketing (segments and positioning) and Requirements (pains → requirements). Note the "confidence level" in the artifact. A human analyst delivers the summary as a monolith: here's what I understood. The agent can attach, to every finding, how many sources it rests on and how well they agree — and pass forward not an opinion, but a picture marked up by reliability.

Inputs: opportunity-brief (phase 1), a map of stakeholder types and pain-bearer types, access to communication channels and calendar.
Tools: meeting scheduler (books calls itself); email and messaging (sends questions, arranges meetings); transcription; insight synthesis; deduplication and pain clustering; a contradiction detector across sources.
Artifact: discovery-report — segments, pains, JTBD, contradictions between stakeholders, a list of open questions, and a confidence level for each finding.
Trigger: a completed opportunity-brief.
Handoff: discovery-report → Marketing (segments and positioning) and Requirements (pains → requirements).

Note the "confidence level" in the artifact. A human analyst delivers the summary as a monolith: here's what I understood. The agent can attach, to every finding, how many sources it rests on and how well they agree — and pass forward not an opinion, but a picture marked up by reliability.

Where it breaks

Trust and consent of a live conversation partner. People talk to an agent differently than they talk to a human. Some will refuse the conversation once they learn there's a model on the other end. Others, conversely, will be more candid without social pressure. But you can't pretend the channel is transparent: consent to talk to an agent, and to be recorded, is a separate question — legal and ethical.

Reading between the lines. The agent takes words in well. What it does worse is what sits behind the words: the power dynamics in the room, what the person is afraid to say out loud, where they're being politely dishonest out of social desirability. An experienced interviewer reads the pause and the topic dodge; the agent, so far, mostly reads text.

Promises and data made in conversation. When the agent promises something to the person it's interviewing, or collects their personal data, someone — a legal subject — is accountable for it. Autonomous outbound contact is not just throughput — it's obligations someone has to hold.

What stays human

Perhaps a live presence at key stakeholder meetings, where trust decides more than information. But even this is a candidate for compression: voice and avatar agents are erasing the line of "is there a human on the other side," and some people are already more candid with a machine than with a live interviewer. What's non-transferable isn't the conversation itself, but accountability for the obligations and data it produces — and that, again, traces back to the principal (phase 14).

human remainder ≈ 86%

Provocation / thesis

Most of discovery is transport and structuring of information, not empathy. Finding the people who feel the pain, asking the right questions, losing nothing, distilling it into structure, marking it up by reliability — all of this is bounded by human throughput, not by the depth of a person's soul. Give the agent a communication channel and a calendar, and the phase's throughput jumps by an order of magnitude, while "genuine human contact" shrinks to a narrow remainder of trust — one that keeps thinning as people grow used to talking to machines.

Running case

Vitrina in this phase

Under the Vitrina hypothesis, the agent decides who to ask: coffee shop, hair salon, and local-store owners as pain bearers; plus a stakeholder — the founder, who has their own vision of the product. The agent writes to fifty small-business owners itself, arranges and books fifteen-minute calls, conducts and transcribes them.

It distills without loss and finds what a human analyst might have smoothed over to fit the original hypothesis: the main pain isn't "I don't have a website." Most of them don't need a website and won't maintain one. The pain is "customers want to order and pay online, and all I have for that is messaging back and forth." So the value isn't the storefront as such — it's capturing the order and the payment. The agent also flags a contradiction: the founder sees the product as a beautiful design builder, while the market sees it as a payment pipe with a minimal storefront. It marks this as a high-importance open question.

Artifact → Vitrina's discovery-report. The "design vs. payment pipe" contradiction moves forward — into Marketing and Requirements — clearly flagged, not lost.

In practice

How it actually works — engineering breakdowns

Standalone howto from practice, showing this phase on real code and a working artifact.

Clarify Gate in Jira: Intake to a Mature TicketClarify gate: task admission by quality of the ask — the same intake an agent runs.
Cross-org Agent Negotiation Through a FirewallTwo agents negotiate a spec through a firewall — autonomous requirements gathering in practice.

Building AI-driven delivery in-house?

Designing the ADLC loop: where the agent holds the role and where the human principal remains — for your team and product.

Email me

The transition engine

Next Move Engine — the system that takes a team to an autonomous delivery loop.

Next Move Engine →