Skip to main content
ciaramaria.xyz
REUSABLE AI DELIVERY SYSTEMS

AI Harness Engineering

Reusable method packs and run patterns that make agent-assisted work scoped, reviewable, and evidence-backed.

AI-native deliveryAgent systems architectureTechnical reviewGovernance & validation

Work completed as part of my role at Whitelabel Co.

WHAT IT IS

The operating layer around agent work

A harness is not the agent and it is not the runtime. It is the method around the work: what evidence to gather, what policies apply, what artifacts must exist, where a human approves the next step, and how readiness is proven afterward.

Harness or method

The reusable instructions, contracts, policies, templates, gates, and scorecards for a class of work.

Agent or runtime

The system that actually executes model calls, tools, sessions, memory, retries, credentials, queues, or schedules.

Human approval

The explicit decision point where a person approves scope, implementation, risk acceptance, or release.

Outcome proof

The artifacts, checks, source links, command output, screenshots, and scorecard that support a readiness claim.

SUITE

Preview the harness portfolio

These are representative examples, not the full set. Some have been used on client builds; others are validated packages or reusable patterns until a target run produces its own evidence.

Greenfield websites and web apps

Greenfield Harness

Guides Figma-first Next.js and Tailwind builds through scope, research, planning, approved implementation, localhost checks, and a scorecard.

Figma sourceNext.jsplanbuild evidencescorecard
Search and answer-engine visibility

SEO/AEO Harness

A portable method pack for auditing routes, canonicals, structured data, content, crawler policy, AI-mediated discovery, measurement, and readiness.

auditstrategyapproved changesindependent verifySCORECARD.md
Brand-system migration

Rebrand Harness

A governed kit for product rebrands with token source-of-truth, evidence-typed values, human gates, and parity checks across product surfaces.

tokenssource evidencehuman gatessurface parityscorecard
RUN FLOW

How a real run proves readiness

  1. 01

    Scope the run

    A human defines the target, goal, access level, owned paths, constraints, and stopping rules.

  2. 02

    Gather evidence

    The agent or engineer collects source material, repo context, live observations, and known risks against the selected method.

  3. 03

    Plan the work

    The run produces a written plan with proposed changes, verification scope, open questions, and approval needs.

  4. 04
    Human gate

    Human approval gate

    A person approves, redirects, narrows, or stops the run before any implementation happens.

  5. 05

    Approved execution

    Agents or engineers make only the authorized changes, inside the allowed paths, with write records.

  6. 06

    Independent checks

    Verification re-fetches, tests, compares, or reviews the result separately from the person or agent that made the change.

  7. 07

    Scorecard

    A durable scorecard records pass, fail, cannot-verify, evidence links, and remaining risks. The final summary is not proof.

  8. 08
    Verified

    Readiness verdict

    The run ends with a readiness class based on evidence, not optimism. Missing capabilities stay visible.

PROOF

What I can claim

  • Verified client-use claim: the Greenfield Harness pattern has been used by multiple engineers and agents on customer-facing Next.js + TypeScript builds; select builds moved from design to production in under two weeks.
  • Verified package claim: the SEO/AEO Harness has source contracts, skills, workflow, fixtures, validators, scorecard semantics, and handoff docs. That proves method-pack integrity, not rankings, index coverage, or AI citations for a target site.
  • Applied Rebrand Harness pattern: used across a multi-page client marketing site and web app dashboard to guide brand-system migration work across real product surfaces.