AgileGypsy is the AI division of Creweng.

Agent harness engineering

The model is the processor; the harness is the operating system.

Agents built with their boundaries in code: tool schemas, permissions and hooks, context and memory, budgets and approval points, all checked before the model acts. A known-answer set from your own questions is re-run after every change, and a deploy fails when the result drops.

Data ingest and graphing

A graph of your data, and the agents that read it.

Documents, records and code modelled as one knowledge graph that your agents share, built the same way from the same data on every run. Agents read a versioned, read-only snapshot, and retrieval is measured on your own questions rather than judged by feel.

Compliance engineering for AI

Compliance compiled into the build, not filed in a document.

Obligations mapped to controls and evidence, the rules an agent runs on kept in one place, and the line it must not cross written as a test that fails the build. Every regulatory line carries its authority tier (settled statute, draft guidance or no authority found) and is analysis, not advice.

Adversarial testing and governance review

Attack it before your customers do.

A threat model per agent, a probe suite run against it, a person working the failures, and a review of identity, mandates, tool scopes and the audit trail. Findings map to the OWASP Top 10 for LLM applications and the agentic Top 10, with severity, a fix and a re-test, on authorised targets only.

Integrations, onboarding and care

Everything around the agent that makes it work.

Content and data onboarding, integrations with the calendar, CRM, helpdesk and email you already run, and MCP tools built to your systems with rate limits and an audit trail. An evaluation retainer re-runs the known-answer set after launch, so quality does not decay unnoticed.

Governed delivery

AgilePM, for work where the scope, the technology and the rules move at once.

Fixed time, cost and quality with variable scope, MoSCoW priorities, change by written request against fixed timeboxes, and a business case re-tested every increment, including the option to stop. It is how the division runs work that needs a plan, timeboxes and quality gates.

The price and the dates hold

Fixed price, quoted after a short scoping call. Time, cost and quality are fixed when the price is agreed; the scope is what flexes, through MoSCoW priorities set with you at the start.

Change by written request

Every change arrives as a written request against a fixed timebox, is logged, and is re-scored before it is accepted.

Risks reviewed every increment

Risks, assumptions, issues and dependencies are scored on probability and impact and reviewed at each increment. Oversight is by exception, so an escalation is a signal rather than a standing meeting.

Gates and separated duties

Each phase ends at a quality gate, and whoever built a component does not review it. Where the division built a system that Industrial AI Assurance later tests, the report says so.

A business case that is re-tested

The reason for the work is checked again at every increment, and stopping is one of the options.

Handed over in your repository

Everything is code and tests in your repository, with the evaluation set beside it: keep the division on a retainer to run it, or take it in-house.