Research and tools
How can enterprises change faster without losing control of their systems, their investments or their ability to learn?
The books, tools and essays are not separate projects. They are parts of one research agenda, published so that others can learn from it, apply it and challenge it.
Where the agenda reaches
- The economics of AI-assisted engineering
- Reliable modification of large, existing software systems
- Quantitative evidence of system performance
- Organisational decision rights and incentives
- Investment governance and benefit realisation
01
Building organisations capable of change
The Art of Creating Self-Organizing Teams
The human and organisational foundations: autonomy, authority, relationships, leadership behaviour and structural constraints.
Open questions
- When does autonomy hold, and when does structure take it back?
- How do authority and relationships shape whether a change survives contact with everyday work?
02
Measuring and governing change
The Predictability Loop and Flow Observatory
The operational foundations: flow, evidence, forecasting, intervention, uncertainty and delivery economics.
Open questions
- When is a delivery signal strong enough to act on?
- What makes a forecast fit for a commitment?
03 · Open research
Coherence debt
The coherence ratchet, an open-source reference tool
AI-assisted development makes code cheaper to produce without making a long-lived system cheaper to change. Individually safe changes accumulate into coherence debt: the structural cost carried by software that still works and is steadily harder to understand, coordinate and evolve. The name has earlier uses; I use it in this narrower, structural sense.
Open questions
- How can a team see structural worsening arrive when nothing has broken?
- How can a team hold a line against new worsening without first funding a cleanup?
- Which structure has a team actually authorised, and how does it reach the coding agents that change the code?
- When is accepted deterioration a decision, with an owner and a review date, rather than drift?
Tools and working papers
- The coherence ratchet
Open-source tool. Measures what a change does to a codebase's structure, puts the structure a team has authorised where coding agents read it, and holds a line against new structural worsening, starting from today's state rather than a cleanup. A deliberately minimal reference implementation under the MIT licence.
In its published experiment, agents asked to change an order total computed in two unrelated modules changed only the one they could read, in every trial on two vendors' models; with the second module's code in reach they got it right every time. The write-up states models, dates, trial counts and the scoring error it found and corrected.
04 · Book in progress
Brownfield AI
Brownfield AI: Shipping Safe Changes to Existing Systems with Coding Agents
Forthcoming from Manning.
Writing
- Brownfield first: why you can’t just let AI rewrite your codebase
- Specification-driven delivery: governing what AI builds
- The delegation problem: why AI made intent the new bottleneck
- AI generates faster than you can review. That's your real constraint.
- Copilot made your velocity number meaningless. Now what?
Test an idea against your system
These are open questions. If you are working on one of them, or want to test an idea against your own delivery system, I would like to hear from you.
Discuss a research questionFollow the research
Occasional emails from Roman about new writing, books, Flow Observatory and community events. Unsubscribe at any time.