Code Intelligence & Code-Graph Indexing for AI Agents

Narrative

{ "what_is_gaining_traction": "The clearest 2025-early-2026 trend is a shift from raw grep/read workflows toward structured context layers: MCP as the transport, SCIP and language-server semantics for precise navigation, and tree-sitter/embedded stores for local or repo-scoped maps. Thoughtworks’ 2025-2026 Radar volumes and Sourcegraph’s 2026 material both point to context engineering as the center of gravity. (thoughtworks.com)", "local_vs_enterprise": "Local/embedded tools are converging on SQLite, tree-sitter, and lightweight graph stores for small-to-mid repos, while enterprise platforms emphasize SCIP, code graphs, and federated retrieval across repos and history. Sourcegraph explicitly positions SCIP as more precise than embeddings or text search, while Kuzu-style systems show how graph+vector hybrids are being used for agent memory and retrieval. (sourcegraph.com)", "semantic_syntactic_embedding_tradeoff": "Semantic tooling wins when exact symbol navigation, references, or refactors matter; syntactic/tree-sitter wins for fast local parsing and repo maps; embeddings win for approximate recall and tool discovery, but are less precise. The strongest 2025-2026 sources frame the winning pattern as hybrid rather than single-modality. (claude.com)", "mcp_bridges": "MCP-to-code-intelligence bridges are becoming the dominant integration pattern: GitHub MCP Server, Sourcegraph MCP, Serena, agent-lsp, and registry efforts all aim to expose code context as callable tools rather than as static files. The maturity varies widely: Sourcegraph and GitHub are production-facing; Serena and agent-lsp are promising but still evolving; several newer tools are best treated as emerging or experimental. (infoq.com)", "evidence_quality": "The most grounded evidence comes from Sourcegraph’s internal cost modeling, Thoughtworks’ practitioner Radar assessments, and systems papers that include implementation details and benchmarks. By contrast, many local code-graph tools are still product pages, GitHub READMEs, or community projects with anecdotal claims rather than large-scale measurements. (sourcegraph.com)", "outlook": "The field is heading toward hybrid context stacks: build-aware or LSP-aware semantic retrieval, tree-sitter-based local indexing, graph/vector retrieval for cross-repo memory, and MCP as the agent-facing protocol layer. The practical selection criterion is likely to be repo size and precision needs: small local repos can favor embedded indexers, while monorepos and enterprise estates will continue to need SCIP- or knowledge-graph-backed infrastructure. (thoughtworks.com)" }