Speakers: John Law (Accesso)
Track: Software Engineering
Our main product is a conversational agent that helps users explore and analyse their data. As LLM providers released capabilities such as finer-grained cache checkpoints and tool search, the framework underneath our agent became a bottleneck. New features took time to reach it, while some capabilities we wanted did not exist at all. To move faster, one engineer built and shipped a new harness that replaced the entire framework in two weeks.
One of our key decisions was to use an append-only event log as the source of truth for agent execution. Recording model responses, tool calls, results and control events gives us a foundation for pausing an agent for human approval and resuming an interrupted loop by replaying its history. The trade-off is additional complexity: the log must be transformed for each LLM provider and the user interface, while strict append semantics are essential for reliable replay.
Coding agents helped us prototype workflows, create regression and parity tests, and conduct adversarial reviews. This covered more ground faster, but the tests still missed behavioural differences and the agents tended to introduce unnecessary abstractions. Moving this quickly created another problem: it became harder to bring the rest of the team along on the design journey. We changed our workflow to prioritise high-level architectural documentation so the system could be understood beyond the person and agents that built it.
I’ll share the event-log architecture, what enabled one engineer to ship it in two weeks, where agentic coding fell short, and how faster implementation changed the way we communicate engineering decisions.
https://data.webdirections.org/ai-engineer-sydney/sessions.json#b31f06c4-eb5d-4f40-a682-c24217a4c0a2
Full text on webdirections.org
.