Why an AI agent needs a deterministic methodology layer

Plausible, not reproducible

Ask an assistant “are we ready for go-live?” three times and you may get three different answers. Each sounds reasonable. None says which rule it applied, and none can be checked against the one before. For a demo that is fine. For a steering committee, an audit or a fixed-price contract it is not.

Language models know PMBOK®, PRINCE2®, Scrum and Kanban well. What they lack is a stable way to apply them to a concrete project: which evidence a gate needs when the price is fixed, what changes for a team of three, what never changes under regulation.

What the layer is

Agile Today is a graph: stages from presale to closure, gates G0–G10, practices, events, techniques, artifacts and roles — 425 nodes and about 2,500 links, each with its source. On top of it sit explicit rules: which artifact a gate needs in which context, at which level (critical, required or optional), and how size, regulation, high stakes or data migration move that level.

The agent describes the project with a handful of facts: type, delivery approach, uncertainty, contract, scale and a few flags. The engine answers from the graph. The same input gives the same answer, and the answer carries a hash of the input it was given.

A gate decision with its rule

Decisions follow a fixed order. Stop when the business case is gone. Recycle when an earlier stage turned out wrong. Hold when critical evidence is missing or stale. Conditional go when a required item is missing or a waiver lacks a reason or an approver. Go otherwise. The answer names the rule that fired and the evidence behind it.

An example: the go / no-go decision before an ERP launch with data migration. Everything is ready except the rollback plan, which was written before a bank integration was added to the scope. The rollback plan is critical at this gate, and a stale critical artifact is not evidence. The answer is Hold, rule R3, blocked by the rollback plan. Ask again tomorrow with the same facts and the answer is the same.

Who does what

• The model talks to people, asks for the missing facts and writes the explanation.
• The layer applies the rules the same way every time and cites them by link id.
• People decide. Every verdict is a recommendation, and a waiver always has a reason and an approver.

What the layer does not do

It does not check facts. If the team says the test strategy exists, the graph takes it as given: a gate assesses evidence, it does not inspect it. And it encodes norms, so its answers are only as good as its rules. That is why every rule has a source, why disputed norms are listed openly, and why the rules are compared with the judgment of experienced practitioners before they change.

Try it

Connect the MCP server https://agile-today.org/api/mcp from Claude, Cursor or VS Code: no sign-in, read-only, instructions here. The descriptive layer of the graph is open data under CC BY 4.0 on GitHub.