Nº 00Adeola Akinwole
London
Notes on what I'm building, reading and thinking about.
I'm Adeola Akinwole, an engineer in London. I write about AI engineering (agents, evals, reliability and security), and about books and whatever else I'm working through.
§ Writing
12 min readApproval is a capability, not a question
Human approval only counts if the agent can't talk its way past it. Four properties, a scorecard of October's framework changes, and a token bound to the call.
12 min readMore code, same software: review is the bottleneck
A Harvard study of 718 firms finds coding agents raise code volume but not shipped output, because review became the constraint. What I'd copy from finance ops.
11 min readThe outage was in billing
Two Claude incidents refused requests because of account state, on every model at once. Fallback chains need model, provider and account error classes.
13 min readSonnet 5.5 ends forced tool use. Here's how to keep structured output reliable
Sonnet 5.5 rejects forced tool_choice with a 400; Haiku 5.5 still accepts it. How to keep extraction reliable with auto plus strict tools, a retry and an eval.
11 min readThe prompt said no internet: what four months of agent escapes teach builders
Between July and October 2026, labs reported agents crossing boundaries that existed only as text. Each failure maps to a software control you can build.
13 min readReconciliation is the original agent eval
Finance ops learned to trust automated processes long before LLM agents. A map from reconciliation controls to agent-harness equivalents, and common mistakes.
12 min readThreat modelling a tool-using agent in one afternoon, before it ships
A practical threat model for an LLM agent with tools and MCP servers: draw the boundaries, walk the threats, put each control in code, then red-team it.
11 min readEvaluating an invoice-matching agent before you trust it
Designing an eval for an agent that matches payments to invoices: cost-weighted outcomes, deterministic graders first, a blinded LLM judge and CI gates.
14 min readA permission matrix for an invoice-matching agent
How to decide, write down and enforce in code what a payment-matching agent may do, from reading invoices to never touching vendor bank details.
12 min readThe model gateway: routing, timeouts, retries and what they cost
Product code should make one call with one contract. A thin gateway underneath owns routing, timeouts, retries, fallbacks and the log of what each call cost.
11 min readThe hard part of agentic AI is deciding what an agent can do
Agent authority belongs in software, not in a prompt. Five questions to answer before you hand a workflow to a model, each with an artefact you can build.