Lightcap Engineering
Engineering notes on Lightcap's agent runtime: failure traces, control invariants, implementation details, tests, and known limits. Written by Faruk Alpay.
- Certifying a conversation before it goes public
A regex scan passed a CV with a phone number on it. What it takes to build a second layer that can actually clear a document, and to earn the word validated.
Published · 9 min read ·
Faruk Alpay - Blocking false completion in agent runs
How a traced crawl failure led to deterministic checks for false completion, goal substitution, and finished-looking residuals.
Published · Updated · 6 min read ·
Faruk Alpay - Enforcing named tools in agent runs
How Lightcap records real attempts, distinguishes search from scraping, and blocks install-only work from passing as a completed operation.
Published · Updated · 6 min read ·
Faruk Alpay - A cheaper delegated build that failed verification
A June 2026 orchestration experiment cut repeated-context cost, but its first crew result did not compile. Here is what the run supports and what it does not.
Published · Updated · 7 min read ·
Faruk Alpay - Why Lightcap replaced nine personas with two compute tiers
How the product moved from character selection to Small and Medium, with two historical one-run measurements and strict limits on what they prove.
Published · Updated · 6 min read ·
Faruk Alpay - Stopping repeated agent work with distinct-progress fingerprints
How typed final states and a set of distinct files, commands, tools, and sources stop repeated work from masquerading as progress.
Published · Updated · 7 min read ·
Faruk Alpay