Component

Nemo — Orchestration

Next-gen codification engine. Plans once with an LLM, serializes the plan to deterministic steps, and replays them without any AI calls on subsequent runs. Faster, cheaper, more reliable.

tinyfish-io/web-agent

Tech Stack

Packagetf-nemo (Python ≥ 3.12)

OrchestrationLangGraph state machine

PlanningGemini 3 Flash + Claude Sonnet

ElementsAgentQL (get_by_prompt → XPath)

StoreDatabricks Lakebase

The Big Idea: Codification

First run: LLM generates a plan — each step is codified into deterministic, XPath-based actions

Serialization: Plan saved as codified_steps.json — a fully deterministic recipe

Subsequent runs: Replay without any LLM calls — just execute the codified steps directly

Validation: Validated plans cached in Nemo Store (Databricks Lakebase) for reuse

10 Step Types

Step	Codification Method
`linear-step`	LLM planning loop (Gemini + Claude)
`click-step`	AgentQL → XPath
`navigate-step`	Self-codifying
`enter-text-step`	AgentQL → XPath
`extract-data-step`	LLM schema inference
`scroll-step`	AgentQL → XPath
`select-option-step`	AgentQL → XPath
`transform-data-step`	LLM code generation
`do-until-step`	Pre-codified only
`if-conditional-step`	Pre-codified only

Nemo Service (FastAPI)

POST /stream-run

Primary SSE endpoint — EVA calls this when routing to Nemo as primary orchestrator.

POST /enqueue-run

Fire-and-forget background run for shadow data collection alongside EVA.

POST /store/lookup

Check Nemo Store (Databricks) for an existing codified plan before deciding the orchestration path.

Code Pointers

web-agent-core/src/nemo/agent.pyWebAgent (main entry point)

web-agent-core/src/web_agent_core/graph.pyLangGraph state machine builder

web-agent-core/src/web_agent_core/stepsAll 10 step types + codifiers

web-agent-core/src/web_agent_core/steps/linearLinearStepCodifier (planning loop)

nemo-service/nemo_service/web_server.pyFastAPI endpoints

nemo-service/nemo_service/nemo_store.pyDatabricks Lakebase client