Harness Node Splitter

Harness Node Splitter

Overview

The horizontal workflow graph says what the agent should do; the harness executes it. Before the harness can run anything, the workflow must be split into nodes with focused responsibilities. The chapter gives one rule that cuts through the design space:

Nodes differ by tool surface, not by prompt.

The sharpest example is security tooling. Kyle Polley's RedAI splits vulnerability discovery across two roles whose distinction is entirely tool-level. The scanner node holds a filesystem and threat-models source code (optimized for recall). The validator node holds a browser driver, an iOS simulator, a network stack, and a scripting runtime, and drives each candidate into a live environment. Swap the prompts and nothing changes — the validator can still drive the browser because it holds the tool. The role lives in the tool surface.

The chapter's Tip operationalizes this:

Before you add a node, list the tools it will call. If the list overlaps more than 80% with an existing node, you have a prompt variation of that node — merge them and vary the prompt. If the tool lists differ substantially, split.

This skill applies that rule to candidate operations and emits each node's constrained context scope — the tool surface it may invoke, the memory slices it reads/writes, and the input/output contract the schema validator holds it to (three of the harness's six surfaces). A tool-less reasoning node (like Example 2-3's classify vs analyze) is distinguished by prompt and DAG position, not tools, so tool-less nodes default to split.

When to Use

• Turning a horizontal-workflow sketch into executable nodes
• Auditing an existing workflow for the common failure mode: one retrieval-and-reasoning node that both queries the graph and does causal analysis, with neither pass at specialist quality
• Deciding whether a proposed new node is a real role or a prompt variation of an existing one (the Tip as a pre-add gate)

Phrases that should invoke this skill: "split this workflow into nodes", "should this be one node or two", "nodes differ by tool surface", "merge or split these nodes", "the 80% tool-overlap rule", "constrain the node's context".

When NOT to Use

• Scheduling nodes into parallel phases. Once nodes exist, ordering them into topological phases with parallelism is investigation-dag-planner (Ch5).
• Deciding vertical vs horizontal. That first-hop routing is dual-graph-router; this skill runs after a request has routed to horizontal/both.
• Selecting which tools a query needs. Filtering a large tool registry for a query is rag-mcp-tool-selection (Ch6). This skill assumes each operation's tool set is already declared.
• Writing the node prompts. This decides the node BOUNDARIES; the prompt content per node is a separate step.

Process

Rationalizations

Red Flags

• A node with a graph-read tool AND a causal-analysis task. This is the retrieval-and-reasoning conflation. Split into a retrieval node (holds the graph-read surface) and a reasoning node (holds none).
• Zero merges on a workflow with obvious prompt variants. Either every operation genuinely has a distinct surface, or the tool lists were declared too finely — check that "get latency" and "get CPU" both declare the same metrics tool.
• Everything merged into one node. The tool lists were declared too coarsely (every op claims "aws"); declare the specific tool each op actually calls.
• A merged node whose prompt variants have incompatible input contracts. Merging is safe only when the variants operate on the same input surface; a contract mismatch means these were different roles after all.
• CLI --help exits non-zero. SKILL.md / CLI mismatch; the multi-harness invariant is broken.

Non-Negotiable Verification

1. Run the benchmark battery.

   python cli.py benchmark
   

   Confirm 9 sample operations collapse to 8 nodes, get_metrics+get_cpu merge (identical tool surface), the RedAI scanner/validator pair stays split (overlap 0.0), and classify/analyze stay separate reasoning nodes.

2. Inspect the split visually.

   python cli.py split
   

   Confirm the [MERGED] node fuses the two metric-query operations and lists both prompt variants, and that the two disjoint execution nodes remain apart.

3. Run the Tip as a pre-add gate.

   python cli.py overlap --a cloudwatch_get_metric_data --b cloudwatch_get_metric_data
   python cli.py overlap --a filesystem --b browser_driver network_stack
   

   The first must verdict merge (1.0), the second split (0.0).

4. JSON scope round-trips.

   python cli.py split --json | python -c "import json,sys; json.load(sys.stdin)"
   

   No exception means the CLI is harness-portable and each node scope serializes.

Security Posture

• Prompt injection. Operation tasks and tool names are author-controlled workflow metadata. This skill only tokenizes tool names and compares sets; it never executes a task or invokes a tool. If operation definitions are ingested from untrusted sources, sanitize tools and task before splitting — a malicious tool name could bias a merge, but cannot execute.
• Least privilege by construction. The whole point of the per-node scope is that a node sees only its own tool surface and its own memory slice, not the full registry or the full vertical graph. Honoring node_scope at execution time is how the harness enforces least privilege (composes with capability-authorization-gate, Ch3, and subgraph access control, Ch8).
• Privilege escalation. No shell invocation, no eval, no dynamic import, no file writes outside the read-only bundled sample-workflow.json. Merging two nodes UNIONS their tool surfaces — review a merged node's combined surface so a merge does not silently grant a node more capability than either input had.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly, by Anthony Alcaraz and Sam Julien), Chapter 2 — Agentic Graph Architecture Foundations:

• "The Horizontal Workflow Graph" — reasoning/execution/decision/validation node responsibilities (Example 2-3).
• "Defining the harness" — the six harness surfaces; the per-node scope this skill emits maps to surfaces 3 (tool registry), 4 (memory interface), and 5 (schema validator).
• "Splitting a workflow into nodes" — "nodes differ by tool surface, not by prompt"; the RedAI scanner-vs-validator worked example; the 80%-tool-overlap Tip.

This Generator-pattern skill runs after dual-graph-router routes a request to the horizontal graph and before investigation-dag-planner schedules the resulting nodes into phases.