# How it works > How Empryo works: a live ranked code graph, an agent loop with parallel sub-agents, and embedded Neovim with layered LSP, AST, and tree-sitter intelligence. Empryo is three things talking to each other: 1. **A live code graph** - every file, symbol, and import, ranked by importance. Updated as you work. 2. **An agent loop** - reads, edits, tests, commits. Can dispatch parallel sub-agents. 3. **Your Neovim** - embedded. The agent edits through the same editor you use. ## The agents | Agent | Runs | Model | |-------|------|-------| | **Forge** | The one talking to you. Orchestrates everything. | Your active model | | **⚡ Spark** | Read-only research (grep, read, analyze). | `taskRouter.spark` | | **🔥 Ember** | Code edits and refactors. Own context window. | `taskRouter.ember` | | **WebSearch** | Multi-step web research with citations. | `taskRouter.webSearch` | Configure which model each one uses in [the task router](/docs/recipes/task-router). The WebSearch agent is the fallback, not the only path. When your model's provider can search server-side, the model searches itself in a single step and no sub-agent is spawned. That is enabled only on measured lanes: Claude first-party, Claude and GPT through the proxy, and OpenRouter's native-search generations. Everywhere else, gateways, local models, unproven lanes, no key at all. The agent above runs and searches through Empryo's own backends. Citations vary by model, not by provider: Claude returns a full source list, while GPT names its sources in prose and reports none structurally. See [`nativeSearch`](/docs/reference/configuration) to turn provider-side search off. ## What the agents share When Forge dispatches multiple sub-agents in parallel, they coordinate through a shared bus: - **File cache** - first reader caches, others reuse. - **Tool cache** - read-only tool results shared across agents and dispatches. - **Edit lock** - concurrent writes to the same file are serialized. - **Findings** - one agent's discovery reaches the others at their next step. Result: N agents don't re-read the same files N times. ## Code intelligence, layered Operations (go-to-definition, rename, diagnostics…) try the best backend first, fall back if it fails: | Tier | Backend | Used for | |------|---------|----------| | 1 | LSP | precise types, workspace rename, diagnostics | | 2 | ts-morph | TypeScript/JavaScript AST ops | | 2 | tree-sitter | 37 languages - outlines, imports | | 3 | regex | universal fallback | LSP runs via Neovim when the editor is open, via standalone servers otherwise. The agent always has it. ## System prompt Every turn, the prompt is assembled from: mode, project info, git context, the Genome, your memory, forbidden files, active skills. The Genome (built by the Genome engine) is personalized per-turn - files you just edited or referenced get boosted. See [the Genome](/docs/concepts/genome) for ranking details, [memory](/docs/tools/memory) for cross-session knowledge, [compound tools](/docs/concepts/compound-tools) for the one-call tools, [code intelligence](/docs/concepts/intelligence) for the backend router.