What happens between your prompt and the answer, and how do you keep an agent accountable?
This is how I explain it. A language model takes your text, breaks it into tokens and turns those tokens into numbers; inside the transformer, attention works out how the tokens relate to each other, and at the end the model picks a likely next token, then does it again. On its own, a model produces text and cannot act on anything outside the conversation. Tool calling lets it ask for actions. The harness, the deterministic software around the model, runs it in loops, hands big tasks to sub-agents and manages its context, and context engineering is how you give a model with general knowledge the specifics of your business, inside a context window that is always limited.
Once you build agentic systems on top of that, the questions are about control. How much of the system is deterministic, and how much needs the model’s probabilistic output? Which decisions can only a human approve? How do you prove what an agent did? That takes evidence, end-to-end traces like the ones we rely on in application performance management, and audit logs. Without them, enterprise-grade reliability and accountability are very hard to achieve.
I learned this hands-on: a transformer written from scratch, small models fine-tuned on Kaggle, a one-off red-teaming exercise in 2025, and now a trial project for an internal use case at work, where I focus on harness engineering and context engineering, and part of what I am testing is which models and sub-agents respond better to a task.
§ 01
Essays
The first essay on this topic is being written. Subscribe to the RSS feed to read it when it is published. Until then, the sections below list what is filed under this topic.
§ 02
Research
Research · Red-Teaming Challenge: OpenAI gpt-oss-20b, Kaggle ·
Narrative and Reference Prompting as Safety Bypass Techniques
Red-teaming OpenAI’s gpt-oss-20b
My writeup for Kaggle’s Red-Teaming Challenge on OpenAI’s open-weight gpt-oss-20b. I tested whether the model’s safety layers could withstand indirect or nuanced prompting, using conversation-based strategies that mirror how real users might bypass safeguards, intentionally or not. With narrative-style dialogues and reference-based prompting, the model was coaxed into producing disallowed information.
Department newsletters, built with GitHub Copilot CLI
There was a need for internal newsletters, monthly and half-yearly, one for each department, and I build them with GitHub Copilot CLI and not Microsoft 365 Copilot.
How do you help colleagues explore where AI becomes the best tool for their work?
Which AI tool fits a piece of work depends on how its harness, environment and context work; when a task works in GitHub Copilot CLI, I turn it into a skill colleagues can run.
How do you make cybersecurity understandable to people who do not work in it?
My view, which I have not put into practice in my organization, is that awareness programs should use a careful, thoughtfully designed skills survey and request training completion by role, skill level and necessity.
How do you market technical products, such as AI, cybersecurity and developer tools?
When marketing cybersecurity products, we need to understand the pain points very clearly, and the solution we propose should clearly show its technical value.