Why Agents Fail Across Repos—and How to Fix It

The real ceiling on AI coding agents isn't model intelligence—it's context fragmentation. Agents see only one repository at a time and forget everything between sessions. Developers become the memory bus, re-explaining changes across every consumer and re-describing past decisions with each new prompt. This inflates token costs and drowns productivity gains.

Polygraph, an agent-agnostic meta harness by Victor Savkin, unifies dependency graphs across all reachable repos (owned and open source) while preserving session state, creating an illusion of single navigable codebase with photographic memory of all interactions, allowing multi-repo changes to be coordinated as a monorepo and seamless session handoff across developers with full agent history.

The immediate implication: the re-explanation tax can be eliminated, and cross-repo agentic workflows become practical. But there are tradeoffs—dependency graph accuracy, CI orchestration complexity, and governance of shared agent traces all require careful management. The approach also introduces an extra layer between developers and agents, which could mask errors or create over-reliance.

For large polyrepo organizations pushing agent adoption, this kind of infrastructure may soon be more valuable than chasing the next model upgrade. It reframes the problem from "better agents" to "better organizational memory for agents," a shift that has implications for tooling budgets, CI/CD architectures, and how teams share development context.