Three teams independently shipped solutions for AI coding agents losing context across multiple repositories

Neilos (March 27), Mabl (April 8), and Meta (April 6) each published how they solved the same problem: AI agents shipping code that breaks consumers in other repositories the agents didn't know existed. Mabl manages 100+ repositories with 39% of commits AI-assisted as of February 2026. Meta built a multi-stage pipeline with 50+ specialised agents.

Mabl and Meta both built queryable dependency graphs as separate infrastructure. Mabl's is an 850-line structured registry that agents consult at planning time; without it, 'context drift dropped from ~40% of our failures to <5%.' Meta's dependency index reduces the token cost of answering 'What depends on X?' from ~6000 tokens (multi-file exploration) to ~200 tokens (single graph lookup). Neilos used a different model at solo scale: a manager agent holding the cross-repo plan in working memory plus a TOML file mapping projects to paths.

The dependency graph as a queryable substrate emerges as a pattern when teams scale beyond solo operators. Mabl and Meta treated it as table stakes. The body notes that at a 50-engineer org this represents a fully-loaded engineer-month of upfront work and ongoing maintenance, and at a 200-engineer org it becomes a dedicated platform team responsibility.