The trigger was mundane and the consequences weren't. Sales reps at Abnormal Security were giving customers wrong answers about data retention and deletion timelines — a compliance-adjacent mistake that creates legal exposure fast. Andy Chen, an engineer at Abnormal, solved it by building what he calls an Enterprise Context Layer: roughly 1,000 lines of Python orchestrating 20 parallel AI agents that ran for two days and produced 6,000 commits and 1,020 files across 11 knowledge domains in a GitHub repository.
The architecture borrows a <a href="/news/2026-03-14-axe-a-12mb-go-binary-for-unix-style-llm-agent-orchestration">task-claiming coordination pattern that lets agents divide work without collisions</a>. Agents run in a Modal cloud sandbox with plain bash access to a GitHub repo, pulling from Abnormal's internal retrieval and Glean's search API. The output isn't a pile of fetched documents. It encodes institutional judgment, flags where sources contradict each other, and attaches inline citations to every claim. Concrete artifacts include end-to-end customer journey maps cross-referenced against Gong sales call recordings and Salesforce case outcomes, detection model lifecycle documents bridging engineering and customer success, and competitor battle cards backed by specific recordings, product capability docs, and deal outcome data.
Chen's sharpest argument is that retrieval and synthesis are different problems entirely. Glean — built by ex-Google search infrastructure engineers and used inside Abnormal — is described as world-class at finding documents. But it can't encode the judgment required for real knowledge work: knowing a compliance answer changes depending on whether a customer falls under FedRAMP or EU regimes, or that a particular PM consistently announces features before they ship. That synthesis layer, Chen argues, can be bootstrapped by a single engineer in days using parallel agent pipelines, rather than requiring dedicated teams, expensive <a href="/news/2026-03-14-captain-yc-w26-launches-automated-rag-platform-for-enterprise-ai-agents">SaaS contracts</a>, or years of ontology-building.
Hacker News discussion pushed the competitive stakes further. One commenter noted that Cursor connected to Slack and GSuite via MCP already outperforms Glean on many Q&A tasks. Another favored the ECL approach over purpose-built platforms like Dotwork, citing vendor lock-in concerns. For Abnormal — where GTM complexity spans overlapping products, regulated customer segments, and knowledge scattered across Gong, Salesforce, Databricks, and Slack — Chen's system is a direct challenge to the SaaS incumbents currently selling knowledge management. How broadly the pattern travels beyond Abnormal is the open question.