AI does not work on its own

The way we build software has changed, but generating code is not the same as delivering a reliable product.

Does this sound familiar? Teams with AI licences learning on their own, isolated prototypes and adoption squeezed into spare time, without a shared method or common standards. It becomes hard to tell what is maintainable, what has been validated and how to repeat what works.

Working with confidence takes three things:

• Human guidance. People steer, make decisions and validate. AI proposes; the team approves.
• A precise definition of what you need. Requirements, business rules and acceptance criteria. Without them, the agent fills the gaps with assumptions.
• Quality control. Tests, code review, architecture and security checks in every increment, not just at the end.

At Taniwa, we work with AI on real projects. Our offer transfers that way of working to your organisation: building with you, not just showing you a demo.

What is Harness Engineering?

It means designing the environment that gives AI agents the context, tools, boundaries and verification mechanisms they need. This environment — the harness — turns individual use of an assistant into a shared engineering process.

It is more than a collection of prompts. It includes repository instructions, architecture conventions, tool access, automated tests and the points where a person must step in.

A complete cycle with human validation

We refine requirements with the agent, break them into small, verifiable tasks and develop incrementally. We combine test-driven development (TDD), architecture review, QA and security with human approval before merging changes.

Traceability from business needs to code

The chain we aim to maintain is straightforward: business rule → test case → automated test → code. This lets us check not only whether something works, but whether it does what was agreed.

The right context for each task

We organise product knowledge, architecture and conventions so that agents receive relevant information. The goal is to reduce ambiguity and unnecessary context, and make their decisions easier to review.

Adapted to your tools

We start with your licences, repositories, IDE and cloud environment. We assess how tools such as Claude, GitHub Copilot or VS Code fit, based on their capabilities and your organisation’s policies. We prioritise transferable assets and avoid unnecessary reliance on a single vendor.

Our approach: learn by building

Together, we select one of your projects, with a manageable scope and visible business impact, and build it from scratch with your team involved from day one.

We adapt it to your needs, architecture, security requirements and standards. Your team learns from the decisions, mistakes and changes involved in a real product, rather than exercises disconnected from their work.

The outcome is twofold: a working product and an internal foundation for repeating the method.

The pilot is not about clearing the entire backlog or promising a product of any size in four weeks. It is about delivering the agreed scope and establishing a way of working that supports its continued development.

What you get

1. A real product with tests and CI/CD

The agreed pilot scope, deployed in your environment, with unit and integration tests and continuous integration and delivery. We define the acceptance criteria with you at the start.

2. A harness tailored to your organisation

Instructions, prompts, skills, contexts, agent configurations and templates that capture your conventions. A reusable asset that serves as a starting point for other projects, adapted to each context.

3. The ability to keep developing the project

Your team practises managing changes, implementing new features and validating results in the same environment. The goal is for them to continue without relying on Taniwa at every step.

4. Training on your own product

Practical sessions, pairing and direct support. Learning by doing: answering questions as they arise and understanding why each decision is made.

Let’s talk about your pilot project

A typical four-week plan

The duration depends on scope, team availability and the access required. This is our starting point:

Week 1 · Definition and environments

• Agree on requirements, business rules and acceptance criteria.
• Configure the AI environment and initial harness around your standards.
• Set up the CI/CD pipeline and deployment environment.
• Define which data and tools agents may use and which actions require approval.

Participants: Taniwa and your team’s designated leads.

Weeks 2 and 3 · Development and calibration

• Build the product incrementally and refine the harness as we learn.
• Apply tests, review and human validation in every cycle.
• Hold weekly demos to check progress and adjust priorities.
• Involve your team in decisions and practical sessions on the actual code.

Participants: Taniwa leads development, with your team’s involvement and validation.

Week 4 · Delivery and handover

• Deploy the agreed version and run a full demo.
• Work in small groups to practise using the harness.
• Support the first changes made by your team.
• Hand over the assets and review next steps.

Participants: Taniwa and your team, who gradually take over day-to-day operation.

From week five onwards, you can arrange additional support for further product development or other projects.

How we assess the outcome

We agree on an evaluation baseline at the start. At the end, we review the following with you:

• Product: acceptance criteria met and agreed scope deployed.
• Quality: automated tests and checks integrated into the pipeline.
• Traceability: links between requirements, tests and code changes.
• Knowledge transfer: documented assets and changes made by your team using the method they have learned.

We do not promise universal productivity multipliers. We look for evidence in your project and your way of working.

What we need from you

• A pilot project with a manageable scope and business value.
• A technical lead or architect and two to four developers, with agreed part-time availability.
• Access to the necessary tools: AI licences, repositories, environments and corporate standards.
• Time each week for refinement, demos, learning and approval of results.

If you are not sure which project to choose, that is where we start.

Indicative pricing

€15,000 + VAT for a four-week pilot with two full-time Taniwa specialists: an AI Lead Architect and a Senior Software Engineer.

The final price is confirmed in a proposal based on scope and team size.

Includes:

• Development of the agreed product scope.
• Harness configuration and calibration.
• Hands-on training and pairing sessions.
• Handover of prompts, skills, configurations and templates created for the pilot.

Does not include: AI licences, cloud infrastructure or repository costs. We use your organisation’s resources. Follow-on support is quoted separately.

Frequently asked questions

Is this a course or a development service?

It combines development and knowledge transfer. We build a real project while your team learns the method. It is neither a prompting course nor a code handover without support.

Do we have to change our tools?

Not necessarily. We review the tools you use and their compatibility with the proposed workflow. Any required changes are agreed before we start.

What about security and private code?

We define permissions, environments and usage policies before granting agents access. Code and data handling must comply with the terms of your approved providers and your organisation’s policies. AI does not replace security review.

Will we be able to continue without Taniwa?

That is the goal of the handover. You keep the assets and work with them during the pilot. If you need further support, we agree on an additional phase.

From AI experiments to a way of working

We start with a conversation about your context and challenges. Then we select a pilot together, define its scope and agree on the proposal before kicking off.

Bring a project. Let’s build the product and the method to keep it moving forward.