Complex Software
Scale the codebase.
Keep the architectural intent.
For senior engineering teams building with AI: architecture, token economics, parallel delivery and the code harness practices to keep you sane.
Past 200,000 lines of code, what keeps the architecture coherent?
The decisions, boundaries and checks around each change matter as much as the code it produces.
Make the architecture part of the work.
As an AI-assisted codebase grows, a single prompt cannot carry every dependency, constraint and past decision. Give each change a clear place in the system.
- Keep boundaries explicit. Define module responsibilities, interfaces and dependency direction.
- Carry decisions forward. Keep architecture records and repository instructions close to the code they govern.
- Check changes against the system. Use focused tests, review and integration checks to catch drift as work lands.
The token-budget seesaw.
How do you use tokens economically while getting enough useful parallel work through the system to make the most of your available capacity?
Work through context selection, task size, model choice, handoffs and verification. Identify which jobs can run independently, which need shared decisions, and where integration or review becomes the bottleneck.
The aim is useful progress within the budget: enough context to do the work well, and enough coordination to bring parallel changes together.
Discuss your development workflow
From a difficult problem to working software
You may have a problem to solve before you have a technical brief. We help define the requirements and work through the design, testing and delivery.
- AI product and delivery strategy
- Define the problem, what success means and a practical delivery plan before choosing the technology.
- Knowledge and retrieval systems
- Build systems that help people find answers and check the source material behind them.
- Agent workflow and governance
- Define what AI agents can do, where people make decisions, and how errors and handovers are handled.
- Software architecture and production readiness
- Design the interfaces, tests and release process needed to put software into operation.
A fireside chat about agentic software engineering.
Bring a hard engineering question for a discussion with Stephen J. McIntyre: architecture at scale, the development harness, token trade-offs and the practicalities of parallel agent work.
Discuss a fireside chatExperience behind the discussion
Stephen has more than 60B tokens of experience, was in the top 1% of ChatGPT users by usage in 2025, and has delivered AI training for government and private clients.
He is currently pursuing a PhD in AI while delivering complex software hands-on. He has published books on leadership and management, given invited talks on communication in project management (2025) and business innovation and entrepreneurship (2026), and is active in the research community across varied disciplines.
One thing I've been enjoying recently is getting into conversations about how people are using AI and offering insights to help them get to the next level in a variety of areas.
So far, my clients come away with a feeling like: “Wow, I thought I was really into AI, but I see you are at least a year ahead, and I see your insights have saved me a lot of time.”
Half-day advanced code harness training.
Are your developers burning through thousands of dollars in tokens each month? That can be a signal to review their code harness practices and training needs.
For senior engineers whose managers want to make AI-assisted development more deliberate, reviewable and repeatable, because they aren't going to reach out for help themselves.
Discuss half-day trainingWork through the harness around the coding agent: the instructions it receives, the context it uses, the work it owns and the checks required before a change is accepted.
- Mindset
- Limitations of the harness and LLM
- Cyber and information security
- Advanced skills and policies
- Advanced goals, progress assessment and recovery from stalls
- Optimising token, CPU and attention economy
Agree the examples, engineering context and intended outcomes before the session.
How we work
Understand
Identify the users, the problem and the constraints.
Design
Agree how the system should work, including decisions that need a person.
Deliver
Build the software or guide the team delivering it.
Verify
Check the result against the agreed requirements and document its operation.
What is getting difficult in your codebase?
Tell us about the system, the engineering team and the problem you want to work through. We can agree the appropriate discussion, training or software scope from there.
Discuss complex softwareFor an independent research engagement around a board or executive decision, explore our AI Strategy & Research Engagement.