AI-driven engineering

From coding assistance to verified delivery loops.

I work on the engineering around the agent: specifications, evaluation, review, and operational control. These are the delivery patterns I use and the public tools I build to make their assumptions inspectable.

01 Specs before prompts
02 Agents inside harnesses
03 Evidence before scale
3 maturity stages
5 loop steps
4 governance tests
1 measured workflow first
Code and contribution

AI Engineering Journey

Public Python tool · design and implementation

What has to change before a team can delegate more?

I built AI Engineering Journey to turn an organisation’s capability assessment into a focused 90-day roadmap. It examines discovery, workflow, evaluation, architecture, platform, governance, and learning, then identifies the capabilities constraining progress.

The interesting engineering choice is to make the evidence visible. A later assessment is compared with the earlier priorities, so a higher self-reported score is distinguishable from a demonstrated improvement. A workflow-readiness check shows where an agentic pipeline would have to stop for human intervention.

This is a decision-support tool. Its scores depend on the assessment supplied; it does not independently audit an organisation or certify autonomous delivery.

Journey

Three stages from autocomplete to agentic-scale delivery.

AI-augmented

Engineers use AI for explanation, refactoring, tests, and local implementation help. Productivity improves, but the process is still centered on individual prompting.

Primary artifact Prompt, code suggestion, local test.
Leader question Are teams learning the tool or changing the workflow?
Core principle

The model proposes. The harness disposes.

Delivery loop

Spec, plan, implement, validate, replan.

01

Constitution

Capture mission, architecture, constraints, coding standards, and non-negotiables.

02

Feature spec

Define outcome, edge cases, acceptance tests, and what must not change.

03

Agent plan

Ask the agent to decompose work before touching files or running tools.

04

Implementation

Delegate bounded tasks with clear permissions, context, and test commands.

05

Validation

Review diff, tests, trace, cost, rollback path, and the next replan.

Governance

Forward the constraint before the capability.

When to use agents

Use autonomy where evidence is inspectable.

Parallel agents can cover more ground, but they also consume more context and compute. The discipline is to name cost, failure modes, and stop conditions before the architecture becomes fashionable.

Clear outcome and inspectable artifacts.
Reliable verifier: tests, traces, or held-out review.
Bounded, reversible steps.
Explicit owner for escalation and release decisions.
Decision table

The delegation question moves from can it code to can we verify it.

Work type
Good first delegation
Required evidence
Refactor with no behavior change
Agent drafts, engineer reviews boundaries.
Tests unchanged, diff readable, performance stable.
Feature slice
Agent implements a bounded spec.
Acceptance tests, trace, screenshots, rollback path.
Architecture change
Agent explores options, humans decide.
Trade-off memo, dependency map, migration plan.
Course signals

The public curriculum is moving toward disciplined agentic delivery.