Back to News
June 3, 2026Blog

Building AI Agents: From Rigid Graphs to Agent Harnesses

Frank WangFrank Wang· CTO & Co-Founder

At actAVA, our mission is to close the gap between AI research and real-world healthcare adoption. To do that, we had to make a foundational architectural decision early: how do we orchestrate AI agents in a domain where workflows are infinitely variable, compliance is non-negotiable, and the underlying models are improving every quarter? There are two competing philosophies in the field today. One is dying. One is compounding.

Building AI Agents: From Rigid Graphs to Agent Harnesses

There are two competing philosophies in the field today. One is dying. One is compounding.

The Two Paradigms

The old way — the way most teams start — is the rigid graph. Think LangChain, LangGraph, and the long tail of frameworks built on the same idea. You define nodes and edges. Node A extracts the patient data, an edge routes to Node B, which checks HCC codes, and you write explicit logic for what happens when Node B fails. It is structured. It is debuggable in theory. And it encodes a dangerous assumption: that the engineer knows the workflow better than the model.

The new way — the actAVA way — is the agent harness. It is prompt-driven and skills-based. Instead of hardcoding workflows in code, we provide a thin, secure execution environment and let the model's reasoning drive what happens — guided by natural language prompts and well-defined skills.

Two Paradigms, One Winner

RIGID GRAPH Node A: Extract Data Node B: HCC Check fail? yes Error Handler edge Edge case… no Node C: Output Every new rule = new code AGENT HARNESS Thin Execution Loop Assemble prompt → Call model → Execute tools → Repeat B The Brain Frontier model — swappable H The Hands Typed, governed tool sandbox S The Session Append-only audit log New capability = prompt update
The rigid graph predicts every path; the agent harness reasons through each one.

The Clearest Proof Point: LangChain vs. Claude Code

The best way to understand the difference is to look at what has already played out in software engineering.

If you have built with LangChain, you know the pattern. You are building a state machine. You predict and hardcode every possible path a workflow might take. The framework gets in the way of the model — good for prototyping, clumsy at production scale, and brittle the moment reality deviates from your graph.[1]

Now look at how Anthropic built Claude Code. Their philosophy is the opposite: bet on the model. Their harness is deliberately thin — essentially a "dumb loop" that assembles the prompt, calls the model, executes whatever tools the model requests, and repeats.[2] There is no graph orchestrating the agent. The agent figures it out.

The most telling technical detail: Claude Code's planning tool is a no-op. It does not actually execute anything. It exists purely to force the model to write down its plan in context — pure context engineering.[3] The intelligence is not in the framework. The intelligence is in the prompt and the model.

LangChain SDK is the graph pattern. Claude Code SDK is the harness pattern. The harness wins — and it wins more decisively every time the underlying model gets better.

How KORA Is Built

We made the same architectural bet for healthcare. actAVA KORA is built by decoupling three things that graph frameworks tangle together:[4]

Component What it is Why it matters
🧠 The Brain The reasoning model — frontier-class, swappable Every model upgrade is a config change, not a rewrite
🤝 The Hands A secure, governed execution sandbox with typed tools 174 tools, all policy-checked and versioned
📋 The Session An append-only log of everything that happened Human-readable, reproducible audit trail for every decision

Our 13 production-ready agent workflows and 174 agent tools are not a tangled mess of Python nodes and edges. They are modular, prompt-driven skills running inside a SOC 2 Type II compliant harness. The flexibility lives in the prompt layer; the guardrails live in the harness. That separation is what let us ship through v4 of the platform in four months.

The Business Impact

This is not an architectural preference. It changes the economics of building and operating AI in healthcare — across five dimensions.

1. DEPLOYMENT SPEED

Authoring Power Shifts to the People Who Know the Work

In the graph world, changing a workflow is a software engineering cycle. New CMS rule? Engineers rewrite the graph, test the edges, and deploy. In the harness world, workflows are prompts and skills. Updating a workflow often means updating an instruction or adding a skill definition — work that clinical operators, compliance leaders, and SMEs can do directly.

New capabilities ship in hours, not sprints. For a healthcare AI buyer, that means new use cases land in weeks instead of quarters.

2. SELF-LEARNING

Continuous Improvement, Not Quarterly Patches

Rigid graphs do not learn — they execute what they are programmed to do, fail the same way every time, and wait for a human to fix the code. The harness pattern enables true improvement loops. That is our GREEN subsystem: every interaction is captured, every failure is analyzed, and every skill is continuously refined against evals. Improvement does not ship quarterly — it streams.

3. FREE UPLIFT

Every Model Upgrade Is Free — and It Compounds

This is the one to emphasize. When you hardcode workflows as graphs, you encode assumptions about what the current model cannot do. When a smarter model lands, your graph holds it back — sometimes it breaks outright and needs a rewrite. In the harness paradigm, the workflow logic lives in the prompt and the model's reasoning. Frontier model upgrade? Config change. Every deployed agent gets smarter overnight.

We are riding the model improvement curve instead of fighting it — and that compounds, every quarter, for the life of the platform.

4. AUDITABILITY

Compliance by Construction, Not by Bolted-On Policy

In healthcare, explainability is non-negotiable. Debugging a LangGraph failure means tracing through layers of framework abstractions. In our decoupled architecture, the session is the audit trail — an append-only log of the exact prompt, the model's reasoning steps, the tools called, and the results returned. Every HCC coding recommendation, every eligibility decision, every chart audit finding is human-readable and reproducible.

Counter-intuitively, the harness pattern is better for regulated environments than rigid graphs — but only if you build the governance layer right. That is why KORA has the RED layer: every tool typed and policy-checked, every skill versioned, every run traced. Adaptability where you want it; rigidity where you need it.

5. MAINTENANCE ECONOMICS

Skills Are Flat. Graphs Accumulate Complexity Quadratically.

Graphs accumulate complexity quadratically. Every edge case adds a node, every node adds state, and every state adds a test. Eventually, you have a system nobody wants to touch. Skills are flat, declarative, and version-controlled. You read a markdown file, you understand the workflow. You edit it, you have shipped a change.

Hours
Time to ship a new clinical workflow update — versus sprints in the graph world
174
Typed, governed agent tools running inside KORA's SOC 2 Type II compliant harness
4 mo.
Time to ship v4 of the KORA platform — made possible by decoupled, prompt-driven architecture
$0
Engineering cost to absorb a frontier model upgrade — every deployed agent gets smarter automatically

Why This Matters Specifically for Healthcare

Healthcare workflows are not a finite set of well-defined paths. Eligibility, prior auth, chart audits, risk adjustment — these are domains with infinite variation. Every payer is different, every plan is different, every patient is different.

Graph-based agents cannot enumerate the branches. You are either building a graph for every variation — economically impossible — or building one that handles 60% of cases and dumps the rest on humans. Neither scales.

The harness pattern handles the long tail naturally, because the agent reasons about each case rather than executing a pre-defined script. That is how we ship horizon agents that handle month-long enrollment workflows and multi-day chart audits — work that a graph-based system would either oversimplify or fail at entirely.

Healthcare Workflow Coverage: Graph vs. Harness

100% 50% 0% Cases Handled Workflow Complexity / Variation Agent Harness — stays above 90% across all complexity Rigid Graph — drops below 60% in the long tail Long tail Graph dumps cases on humans here Harness reasons through
Graph-based agents plateau around 60% case coverage. Agent harnesses reason through the long tail.

The Punchline

The graph paradigm assumes the engineer knows the workflow better than the model. That assumption was reasonable in 2023. It is not true today, and it gets less true with every model release.

The agent harness paradigm assumes the model is the smartest part of the system, and the engineer's job is to give it tools, expertise, and guardrails — not to micromanage its decisions.

We built KORA on this bet. It is why we deploy faster, improve continuously, inherit every model upgrade for free, and stay compliant by construction. It is the same architectural bet Anthropic made with Claude Code — and it is the bet healthcare AI needs to make to actually scale.

The era of hardcoding AI workflows is ending. The harness outlasts the model.

References & Further Reading

  1. Moore IQ. "Is LangChain Still Relevant in 2026? The Honest Answer for AI Agent Developers." BSWEN, April 2026. On graph-based orchestration brittleness at production scale.
  2. "Claude Code Agent Harness: Architecture Breakdown." WaveSpeed Blog, 2026. Confirms the "dumb loop" pattern: Claude Code's runtime is a thin, repeating cycle with intelligence residing in the model, not the orchestration layer.
  3. Pankaj Pandey. "Claude Code's planning tool is a no-op. LangChain built a framework around that insight." Medium, April 3, 2026. Confirms the planning tool calls write context but execute nothing.
  4. Anthropic Engineering. "Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems." arXiv, April 2026. Systematic analysis of the decoupled brain/hands/session architecture pattern.
  5. Augment Code. "Anthropic Agent SDK: What It Ships vs. What It Leaves to You." 2026. Comparative analysis of Anthropic SDK vs. LangChain harness approaches.