Is this a licence to let agents commit to production?

No. A person approves the deploy. The harness and the guardrails make that review fast enough to be real rather than a rubber stamp.

Do we need this for our first agent?

No. You need it before the third. The cost of retrofitting scaffolds, budgets, and shadow mode across a handful of bespoke agents is the reason programmes stall.

How long does shadow mode run?

Until the proposals match what the desk would have done, on a sample the business owner accepts. We report the disagreements, not a score we invented.

Will you publish a productivity multiple?

No. We will show you the specification, the harness, and the review cadence. Anyone quoting a multiple for your estate has not read it.

Guide · spec-driven engineering

Write the specification. Let the line build the part.

The first week is in the org, reading what already exists. One agent is a project. Ten agents is an operating model: specifications, guardrails, shadow mode, and a review cadence. Engineers stop hand-writing every endpoint and start owning the line that produces them. The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.

Get the written assessment

Certified Partner since 2010 · MVP Hall of Fame · 200+ agents in production · UAE and US desks

The five planes

The shift

From handcrafted code to a line you designed.

This is not “the model writes the app”. It is the same discipline a plant engineer uses: specify, constrain, verify, then run the line.

01
Specify
The desk, the inputs, the allowed tools, the write, and the rollback. If it cannot be written down, it is not ready to be automated.
02
Constrain
A policy catalogue and a scaffold that already carries identity, logging, and budgets. New agents inherit the rules instead of re-arguing them.
03
Verify
A harness that runs before anything ships: schema checks on structured output, policy checks on tool use, tests on the deterministic parts.
04
Approve
A person signs the deploy. The point of the line is faster honest work, not unattended production changes.
05
Operate
Spend, refusals, and behaviour reviewed on a cadence, with an owner per agent.

Spec-driven delivery. The engineer authors the specification and the guardrails. Agents implement against them, the harness verifies, and a person approves the deploy.

Mermaid source

flowchart TD
  E["Engineer: specification, guardrails, policy catalogue"] --> A["Agents implement against the spec"]
  A --> V["Validation harness: tests, schema checks, policy checks"]
  V -->|"fails"| A
  V -->|"passes"| H["Human review and approval"]
  H --> D["Verified deployment"]
  D --> O["Fleet operations: spend, refusals, anomalies"]

Principles

Five rules we hold to.

01The specification is the artefactProse that names inputs, tools, writes, and rollback. It is reviewable, diffable, and it outlives whoever wrote the prompt.

02Scaffold, do not bespokeA new agent starts from a template with identity, structured errors, telemetry, and budgets already wired. Bespoke agent number four is where governance debt starts.

03Shadow mode before write authorityThe agent proposes, a person decides, and we compare proposals with what the desk actually did. Write access is earned on evidence.

04Deterministic backboneModel reasoning sits inside a state machine that can retry, compensate, and replay. The model is a step, not the orchestrator.05A human gate on the expensive hopsInterrupt the graph, ask for approval on the surface the approver already uses, then continue. That is a design pattern, not a fallback.

The team

What people do instead.

Nobody is removed. The work moves up a level, and the review gets sharper.

Questions

What we actually say.

Is this a licence to let agents commit to production?: No. A person approves the deploy. The harness and the guardrails make that review fast enough to be real rather than a rubber stamp.
Do we need this for our first agent?: No. You need it before the third. The cost of retrofitting scaffolds, budgets, and shadow mode across a handful of bespoke agents is the reason programmes stall.
How long does shadow mode run?: Until the proposals match what the desk would have done, on a sample the business owner accepts. We report the disagreements, not a score we invented.
Will you publish a productivity multiple?: No. We will show you the specification, the harness, and the review cadence. Anyone quoting a multiple for your estate has not read it.

The brief · one email

Send one process you keep rebuilding.

We will tell you whether a specification and a harness would have stopped that.