The three phases

Platform is a department at Plain Concepts, and this section is how it builds and runs software, written down. We write it down because "agents merge to production with no engineer in the loop" is the kind of claim that should be checkable rather than taken on faith, and because every Platform project is measured against it. If you are here to set up a project rather than to read about one, go straight to Adopt in your project.

Most Platform apps pass through three phases. They are not a project schedule. What separates them is who is allowed to start a change, and each phase hands that right to a wider group.

flowchart LR
  D["<b>1. Discovery</b><br/>departments and engineers<br/>decide what to build"]
  F["<b>2. Foundations</b><br/>engineers and agents build<br/>the app and its guardrails"]
  M["<b>3. Maintenance mode</b><br/>the business files requests,<br/>agents implement them"]
  D --> F --> M
  M -.->|"too large for a small change"| D
  classDef phase fill:#eef0ff,stroke:#554cff,stroke-width:2px,color:#172033
  class D,F,M phase

The four pillars

Four repositories carry all of this. Foundations, agent-harness and project-health are private repositories and agentic-workflows is internal, visible to everyone in the organisation; the npm packages they publish are public. When a page here is not specific enough and you have access, the source is the real reference.

Taken in order they form a chain, which is the clearest way to see why there are four and not one: Foundations sets the conventions, agent-harness puts them into a repository as something agents can act on, agentic-workflows runs the agents against it, and project-health measures all three. It checks that the guardrails Foundations defines guard each default branch, that the harness and the workflows are installed and current, and how much of what the agents produce merges without a person. Drop any of the first three and the phase it carries stops working. Drop the fourth and nobody can tell whether it did.

What Health measures

Health is not a phase 3 report. Each of its four areas judges a different link in the chain above, and the phase that link belongs to.

One rule runs through all four: a missing metric is a fact about the data, not a gap to fill with a plausible number. Two DORA metrics are refused rather than estimated, a practice Health could only detect is marked as inferred, and a partial total says how much of it was measured.

Phase 2 exists to make phase 3 safe

The claim worth paying attention to is not that agents write code. It is that in phase 3 someone who has never opened the repository can file a request and have it implemented, reviewed and merged with no engineer in the loop.

That is only acceptable because of what phase 2 built. Layering rules that fail the build when crossed. A specification that has to be updated when behaviour changes. A test-quality gate that rejects tests asserting nothing. An audit trail no code path can skip. Each one is a constraint an agent cannot argue its way past, and together they are what buys the autonomy.

So the order is not a preference:

This is the same finding DORA published in 2025: AI amplifies what a team already has, good practices and bad ones alike. See the 2025 State of AI-assisted Software Development.

Outcomes belong to the business, means belong to engineering

Phase 3 works because that line is written down and enforced by the intake itself, not renegotiated for every request.

Anyone can ask for a change to an outcome: how a screen behaves, what a workflow does, what the wording says, what a business rule or a formula should compute. Requests about means stay with engineering: architecture, infrastructure, deployment, authentication, data models, APIs.

The line is drawn by category rather than by size. A small, clear, well-tested request to change how the app authenticates is still an engineering decision, so triage hands it to a maintainer instead of letting an agent build it.

The phases overlap

A project is rarely in a single phase. The same app can be in maintenance mode for its screens while a new integration is still being worked out in discovery, and the foundations for that integration get built while everything else keeps shipping. Discovery in particular never really closes: it runs alongside delivery rather than in front of it.

What does not vary is the sequence for any one capability. A capability reaches maintenance mode only after its guardrails exist.

The phases in detail