THOUGHT · AI-ASSISTED ENGINEERING
Why LLMs can seem so dumb to experienced developers
Why probabilistic models need deterministic engineering around them.
Read the thoughtWriting
Ideas tested against the work.
Architecture, AI-assisted engineering and the systems behind reliable delivery. Published pieces sit alongside the series I am working through next, so the shape of the thinking is visible without pretending unfinished ideas are finished articles.
Published
Available now.
Series catalogue
What I am working through next.
These are approved series-level contracts, not placeholder article lists. Individual article titles are deliberately deferred until each series is ready to be derived from its thesis and learning outcomes.
Token Economy
Tokens are an economic resource in agentic systems. Good architecture spends them deliberately and substitutes retrieval, deterministic computation and reuse where fresh model inference adds little value.
Covers: Understand the major sources of token consumption. Distinguish useful model reasoning from avoidable token waste. Reason about bounded context, retrieval, reuse and deterministic operations as efficiency mechanisms.
Multi-Agent Execution
Multiple agents are useful when decomposition, specialisation or concurrency creates more value than coordination costs. The series treats multi-agent architecture as distributed execution rather than a collection of chatbots.
Covers: Understand task decomposition and specialisation. Compare sequential and parallel execution. Reason about dependencies, orchestration, hand-offs and coordination overhead.
LLM Parity
Agent capability should not be confused with model identity. Robust systems define capability contracts and test where model substitution is safe and where behavioural differences matter.
Covers: Separate agent architecture from model and provider choice. Understand behavioural and capability variance between models. Use capability contracts and parity evaluation to reason about substitution.
Agent Building & Harnesses
A production agent is an engineered system of instructions, capabilities, context, state, tools, controls, observability and verification. The harness around the model is often more important to reliability than the prompt itself.
Covers: Understand the components of a production agent. Distinguish prompts from complete agent systems. Reason about capability boundaries, state, context, observability and execution controls.
Agent QA
Probabilistic systems still require deterministic engineering discipline. Quality comes from observable contracts around uncertain behaviour, deliberate failure testing and evidence before work is declared complete.
Covers: Understand why conventional unit testing remains necessary but insufficient. Distinguish deterministic assertions from probabilistic evaluation. Design regression, failure and acceptance testing around observable evidence.