Andrew Wilson is a London-based solutions architect on Anthropic’s Applied AI team, helping businesses deploy advanced AI systems. His focus is long-running autonomous agents: software that can work for hours without losing context, abandoning unfinished tasks, or confusing an impressive interface with a functioning application.
Wilson studied industrial engineering at Northeastern University and worked in technology consulting and data-product engineering, building analytical models, automating reporting, and coaching startups through the university’s IDEA venture accelerator. His software projects include machine-learning experiments, Kaggle competition code, web applications, and nonprofit leadership-development tools. He subsequently joined Anthropic as a solutions architect in London.
At Anthropic, Wilson approaches agent reliability through three practical failure modes: degraded context, weak planning, and unreliable self-evaluation. His AI Engineer Europe workshop, co-presented with Ash Prabaker, maps those problems onto several engineering principles:
- Agent harnesses must evolve with model capabilities. Permissions, checkpoints, sub-agents, context handling, and verification loops should address demonstrable weaknesses; safeguards can be simplified when stronger models make them unnecessary.
- Persistent artifacts preserve continuity. Feature lists, progress files, initialization scripts, test results, and Git commits let successive sessions recover their bearings, implement unfinished work, and leave an auditable record.
- Context strategy depends on the task. Fresh sessions can reduce drift, while larger context windows and compaction can sustain continuous work; evaluations should determine which configuration performs better.
- Working software requires behavioral verification. Smoke tests, browser interaction, and explicit completion criteria expose features that appear finished but fail in practice. Existing production codebases also demand workflows adapted to their established stacks and quality standards.
- Model selection shapes agent architecture. Stronger models can handle planning while more economical models execute implementation, provided each configuration is evaluated against its assigned task.
Wilson treats Claude Code as a practical environment for testing these patterns and the Claude Agent SDK as a foundation for deploying longer-running applications.
