Our approach

A method for systems that act.

Our methodology starts with a measurable business problem and ends with a system that has explicit requirements, explicit authority, named oversight, and evidence for every claim made about it. It is written down as a sixteen-section blueprint, and the same blueprint drives our consulting work and the Design Studio.

Four commitments

What the method insists on.

Ask whether an agent is needed

Every design starts at the left of the autonomy spectrum. A fixed workflow with one classification step often beats an agent on cost, latency and auditability. We move right only when the evidence says the simpler design cannot meet the bar.

Enforce boundaries in systems, not prompts

Instructions shape behavior; they do not bound it. Identity, scopes, spend limits, allowed tools and irreversibility checks are enforced by deterministic code before any effect reaches the world. The test: which invariants still hold if the system prompt is deleted?

Measure accepted outcomes

Benchmark averages describe a model. They do not describe your system on your cases. We measure the outcomes people accepted, the ones they corrected, and what each cost in tokens, retries and attention.

Earn autonomy through evidence

Each class of action starts at a justified level with written promotion and demotion criteria. Stopping authority has a named owner and a tested path. Autonomy moves up on evidence and moves down on incident, automatically.

The autonomy ladder

Autonomy is a dial, set per action.

A single system usually sits at different levels for different actions: A3 for drafting, A1 for anything that moves money. The blueprint records a starting and a target level for each, and what has to be true to move between them.

A0Inform
The system reads and summarizes. It takes no action and changes no record.
A1Propose
It drafts actions. A person approves each one before anything happens.
A2Act within bounds
Low-risk, reversible actions run inside enforced policy. Everything else still waits for approval.
A3Act and report
It acts across its scope and reports after the fact. Irreversible actions keep a gate.
A4Operate
It runs the process end to end. People own the goals, the limits and the stop.

What we measure

Evidence that predicts the next thousand cases.

Consistency: Does the system give the same accepted answer across repeated runs of the same case? One good run proves nothing.
Approval load: How many decisions reach a person per hundred cases, and how many of those are rubber stamps? Oversight that nobody reads is not oversight.
Cost per accepted outcome: Tokens, retries, latency and human time divided by the outcomes that were actually accepted, not the ones that were produced.
Evidence for promotion: The specific test results, trace reviews and incident-free interval that justify moving an action class up a level.

“The model proposes. Deterministic software decides whether the proposal is allowed, executes it if so, and records what happened.”

From the manuscript The Agentic AI System Design, the source of this method. Design Studio is its application in software.