05. September 2026

Agentic Development, Part 4: Intent Harvesting and Pairbuilding

Mikoláš Aleš's illustration of the Maharal of Prague and the Golem

This post is about addressing the new bottleneck of software: intent clarification. We can also call it specification or alignment. In the dark factory post, I described what happens when agents can handle most of the production. The question now is how we collect and clarify the intent that will guide that production. I think of this as intent harvesting.

We have passed another step in AI transformation. Agents are now part of the living context. We are moving from a vertical organization of humans, AI, and context, in a 1-to-many relationship, toward groups of humans and AI working together in a many-to-many relationship.

The first time I worked in a common group of humans and AI, I had the same positive feeling as when I discovered pair programming and the level of creativity it unlocked at that time. I had that feeling again, with the same intensity, and this time I felt I was connecting with more people than only developers. I was pairbuilding with business people in a live Slack channel.

AI killed pair programming, but it has created pairbuilding.

The idea is pretty simple, and it has several components. Instead of bringing the context to AI, we should bring AI to the context. The context of the company is already digital, and this is good news. The challenge is to capture it where it lives.

The other good news is that most companies are used to remote work. The better a team knows how to work remotely and asynchronously, the easier the agents' context harvesting mission becomes. The agents can participate where people are already discussing the work, asking questions, and clarifying what they want to achieve.

The process of building is radically changing and is increasingly based on prototyping on the business analyst's side. The BA formulates an intent, usually after a long interaction with AI. Sometimes the AI itself suggests a radical improvement to an existing process. Together, they make a first artifact that we can consider a seed.

This seed is submitted in Slack. Based on the impact and size of the requirements, it can be processed directly by AI, as with bugs and reports, or challenged with the relevant contributors. Bigger requirements are added to the general roadmap following clarification.

During this process, the AI is public in the channel, challenging the specification and pinging coworkers to make sure nothing is missing. Everyone contributes to the same discussion. We have gone through that clarification in under an hour, for a process that would previously have taken days.

At the end, the quality is much higher than what we used to obtain with human specifications. The result usually contains a small prototype that will be used in the next step for design validation and workflow inference.

The dashboard example below shows how people and AI share the work. Names have been replaced with roles.

How a dashboard gets builtOne shared Slack thread
  1. ScopeRequester
    Posts the specification and mockup; records the request.Slack · Asana
  2. AssignEngineering lead
    Routes to metric owners and tags the AI agent on the pull request.Slack · GitHub
  3. ReviewAI agent
    Reads the code, runs tests, and checks numbers against the data.GitHub · Toolbox
  4. Split the workEngineering lead
    Assigns plumbing, architecture, and security to AI; metric definitions to their owners.Slack
  5. DecideAI + metric owners
    AI turns each disagreement into a question with one owner. Owners answer in the thread.Human gate 1 · Define what each metric means
  6. BuildAI + reviewers
    AI commits to a branch and opens a pull request. Automated checks run; people review.GitHub Actions · dbtHuman gate 2 · Approve the changeHuman gate 3 · Merge to dev
  7. PromoteEngineering lead
    Opens the dev-to-main promotion request. The deployment pipeline runs.GitHub → Cloud RunHuman gate 4 · Promote to production
  8. VerifyAI agent
    Checks the deployment and queries production to document each number.GitHub · BigQuery
Follow the steps from top to bottom. AI and people share the work; the four human gates remain with people.

After that, most of the work is done by agents behind the scenes, following the strict process I described in the dark factory post. During this work, engineers intervene to check quality and ensure the architectural backbone is safe. The agents notify the requester and stakeholders and ask for their validation before requesting promotion and delivery to production.

The power comes from keeping the context in one place and making sure AI has safe and complete access to the relevant sources and tools. Together, these make the whole process shrink from days to hours.

This transformation comes, of course, with constraints and requirements. But I feel most remote, asynchronous teams are already used to them: documentation and explicit communication, even when that can be considered verbosity; transparency and clean context, keeping work discussions out of private conversations and avoiding mixing topics.

These are the conditions that let agents capture the context where it lives, and let us clarify intent together.

Cover: Mikoláš Aleš, The Maharal of Prague and the Golem, 1899. Public domain image via Wikimedia Commons.