The notebook
Things I'm learning in public.
Experiments when I have data. Reflections when I have experience. Thoughts when I'm still working something out.
Are specialised agents just one harness wearing different faces?
Talk of specialised agents left me with the uncomfortable feeling that I had missed an entire layer of AI design. I went looking and mostly found the same general model in a fresh context, wearing a role supplied by a prompt. Useful perhaps, but closer to a mask than a specialist. The operating boundary I expected remained an idea to test.
Read the reflectionThe same retrieval map saved tokens for one AI—and added them for another
The same reference-map layer reduced cost and improved recall for Sol, but increased token use for Luna.
Read the experimentI kept rebuilding document retrieval
Cleaner mission context made retrieval work feel easier to inspect. Turning that observation into a reusable library created somewhere the improvement could be tested.
Read the reflectionAI made me so productive my work is out of date before I finish it
AI shortened my learning loop until I could refine, improve and prove a better approach before the original work was finished.
Read the reflectionI lost control of my AI system. So I built a second one.
When an AI-native system resisted corrections to its own map, I started a controlled parallel experiment around visible seams and human authority.
Read the experimentThe logo was not a tracing problem
A complex AI-generated logo became usable when I separated the work that needed editable geometry from the work that needed painterly cohesion.
Read the experimentThe aquarium outgrew its map
Two days into building an AI-native ecosystem, the implementation was moving faster than its own account of what was true.
Read the experimentAn aquarium for AI
Not software built by humans, for humans, with AI running it. A living ecosystem designed for AI—with one pane of glass left for me.
Read the experimentI made the findings clearer. The repairs got worse.
A failed attempt to improve AI-assisted code repair by restructuring the findings that entered it.
Read the experimentWhen ‘resolved’ only meant well formed
Three agents repaired the same seven findings and all declared success. Independent evaluation found only one repair was complete.
Read the experimentThe best AI reviewer still missed three defects
Three independent reviewers found seven defects between them. The strongest found four—and missed three that mattered.
Read the experimentStart at the seam: a low-disruption way to introduce AI
A worked example of fitting AI into an existing software delivery process through advisory PR review, staged testing and evidence-led expansion.
Read the reflectionDon't onboard a software team to AI. Onboard AI to the team.
A systems approach to introducing AI: discover how work really happens, align competing objectives and prove value through small, measurable changes.
Read the reflectionThe validation race that crippled my PR workflow
Testing every pull request against the latest main created an unfair race. Evidence should become stale when its assumptions change, not merely when the repository moves.
Read the reflectionI measured the wrong context—and the system carried the cost forward
Temporary working context and the context carried by a continuing mission do not impose the same burden. A retrieval experiment changed what I measure and why.
Read the experimentWhat if the agent is not the unit of work?
After queues of 30 to 150 pull requests, a theory: persistent work and transient executors may sustain more progress than bursts of parallel output.
Read the reflectionWhat changes when behaviour meets a cheaper model?
The first behaviour-first experiment showed that complete requirements mattered. The second asked whether behavioural structure could help cheaper coding models keep up.
Read the experimentWhat changes when behaviour comes before the code?
A post about behaviour-first coding sounded promising—and oddly familiar. I built a pilot to find out what the model, the detail and the earlier thinking each changed.
Read the experimentThe case of the review bottleneck
How AI-assisted delivery moved the constraint into assurance—and a phased, composable plan for investigating it.
Read the reflectionWhen prompts grow up: From text files to skills—and the system that followed
How saved prompts became independent skills, dependable chains and an orchestrated AI system governed by evidence rather than instinct.
Read the reflectionTesting databases for AI documentation retrieval
A lexical index, a vector store and a typed graph, each measured on what it improved, what it cost and what it left unchanged.
Read the experimentFree your skills. Let the models compete.
Shared intent can make AI capabilities portable without forcing every provider through the same harness.
Read the reflectionBe kind to yourself. Kill your heroes.
When we judge an AI agent by whether it completed the task, its improvisation can become an invisible dependency.
Read the reflectionAI should learn to give us the stare
The most useful assistant may not answer immediately. It may make the problem, evidence and assumptions visible before helping solve it.
Read the reflectionWhat the green light left out
A system can store the right result and still fail the next decision when its readers cannot see what the result means.
Read the reflectionThe safeguard that made every change wait
What a one-to-three-hour validation suite taught me about matching assurance to the rate and risk of change.
Read the reflectionWhen AI works from an unfinished map
Fast parallel output can hide unresolved shared decisions, letting dependent work move quickly with different assumptions.
Read the reflectionWhen AI finds the answer that used to be right
A train-time thought experiment about what happens when AI retrieves an answer that was once accurate but no longer governs the present.
Read the thoughtSometimes the breadcrumbs are the story
AI can polish away the questions, evidence and rejected paths that give a conclusion its meaning.
Read the thoughtStable deployment slots for agent environments
Agentic workflows expose systems built around durable identities and serial work. Stable slots are a first step towards safe, bounded parallelism.
Read the thoughtWhen humans hold weak systems together
When repeated work depends on memory and vigilance, the system has made human adaptability its operating model.
Read the thoughtWhen AI velocity outruns feedback
AI can accelerate production beyond our ability to understand, verify and correct what it creates.
Read the thoughtAI multiplies output, not human capacity
AI can multiply production faster than accountable human understanding can absorb it.
Read the reflectionThe context compression frontier
What happened when token use, retrieval depth, latency and evidence quality stopped moving together.
Read the experiment