Agent Engineering Lab
A working publication on building agent systems that survive contact with reality.
Agent Engineering Lab is where the practice of building production agents gets written down, argued about, and tested. The premise is plain: most agents are workflows in costume, and the reliability that matters lives in the harness, the loop, and the gate, not in the model. The Lab takes that seriously and earns the complexity.
The streams
Four ongoing streams, each with a different job.
- Field Notes are observations from inside the work, grounded in recent papers, tools, and incidents.
- Signals read the discourse around agent engineering and say what to keep and what to skip.
- Recipes are buildable-in-an-afternoon walkthroughs with verified, runnable code.
- Lab Notes are empirical experiments with measured comparisons and honest caveats.
Behind them sits the reference shelf: Evidence (the measured data behind the claims), Patterns (reusable architectures named across the work), and Projects (teardowns of reference systems).
The Bench
The Bench is the Lab's playground: interactive, runnable experiments that let you manipulate the concepts the pieces argue, in the browser, with no setup. Read the Field Note, then go break it on the Bench.
Cadence
Field Notes twice a week, Recipes on Wednesdays, Lab Notes every two to three weeks, Signals as the discourse warrants.
Who runs it
The Lab is written by Sunil Prakash, an engineering leader and researcher working on enterprise AI systems, agent identity, and production agent architecture. The book Agentic AI for Serious Engineers is the canonical reference the streams draw on, and the Agent Identity Protocol is the related work on verifiable, delegable agent authority. Both are companions to the Lab, not its parent.
Read along
New work lands here first. Follow on Substack, the LinkedIn newsletter, or the RSS feed.