Harness Wars — OpenAI vs Anthropic, and What We Need to Prepare For

0. 30-second summary

• In March–April 2026, Anthropic and OpenAI raised "harness" as an official topic of discussion about three weeks apart.
• Anthropic — 2026.3.24, "Designing harnesses for long-running applications" (Prithvi Rajasekaran)
• OpenAI — 2026.4.15, Agents SDK 2.0 — long-horizon harness, sandbox, subagents

• MCP, skills, and CLIs are being pushed as standardized and open — but for the harness layer, both vendors are signaling "design it inside our SDK."

• At the same time, infra costs for both PCs and local inference are exploding. Even though Google's TurboQuant (2026.3.25) cuts memory by ~6×, the Jevons Paradox means total demand goes up, not down.

• We have to pick one of two paths — play on top of a vendor-locked harness, or design a hybrid native harness ourselves.

1. Why is everyone suddenly talking about "the harness"?

"Claude Code serves as the agentic harness around Claude: it provides the tools, context management, and execution environment that turn a language model into a capable coding agent."

— Anthropic official docs (Memorizer: How Claude Code Works — Harness, AI Agent, Agentic Loop)

flowchart LR
    Model["LLM<br/>(the horse = inference engine)"]
    subgraph Harness["Harness (the tack)"]
        Tools["Tools<br/>Bash / Read / Write"]
        Perm["Permissions &<br/>approval gates"]
        Sandbox["Sandbox"]
        Session["Session &<br/>memory"]
        Context["Context window<br/>management"]
        Ext["Extensions:<br/>MCP / Skills / Hooks"]
    end
    Model --> Harness
    Harness --> Agent["AI Agent<br/>(judges & acts on its own)"]

2. Anthropic's harness — "Start simple, add complexity only when needed"

2.1 Design philosophy

"Find the simplest solution possible, and only increase complexity when needed."

— Prithvi Rajasekaran, Anthropic Labs

"Every component in a harness encodes an assumption about what the model can't do on its own."

2.2 Core pattern — Generator / Evaluator / Planner (GAN-inspired)

flowchart LR
    User["User<br/>(one-line prompt)"] --> Planner["Planner<br/>(turns prompt into detailed PRD)"]
    Planner --> Generator["Generator<br/>(actual builder)"]
    Generator -->|"Sprint Contract<br/>(agreed definition of done)"| Evaluator["Evaluator<br/>(QA via Playwright MCP)"]
    Evaluator -->|"FAIL"| Generator
    Evaluator -->|"PASS"| Done["Done"]

"The generator and evaluator negotiated a sprint contract: agreeing on what 'done' looked like for that chunk of work before any code was written."

2.3 Two findings Anthropic emphasized

2.4 Cost data (real numbers)

3. OpenAI's harness — "long-horizon harness × sandbox × 100+ LLMs"

• <https://openai.com/index/the-next-evolution-of-the-agents-sdk/> (2026.4.15)

• TechCrunch, "OpenAI updates its Agents SDK..." (2026.4.15)

3.1 Key concepts

"The harness refers to the other components of an agent besides the model it's running on, and an in-distribution harness allows companies to both deploy and test agents running on frontier models."

— TechCrunch (2026.4.15)

3.2 OpenAI's message

• MCP is "open" — but the harness is meant to be assembled visually inside OpenAI's workbench.

• In return you get ChatKit / Widget Builder / Connector Registry as one bundle.

4. OpenAI vs Anthropic — harness comparison

flowchart TB
    subgraph Anthropic["Anthropic Harness — code-first minimalism"]
        A1["Claude Agent SDK"]
        A2["Skills + Hooks"]
        A3["Sub-agent pattern<br/>Planner → Generator → Evaluator"]
        A1 --> A2 --> A3
    end
    subgraph OpenAI["OpenAI Harness — full-stack workbench"]
        O1["Agents SDK 2.0<br/>(long-horizon harness)"]
        O2["AgentKit / Agent Builder<br/>(visual canvas)"]
        O3["Sandbox + Subagents<br/>+ Code Mode"]
        O4["ChatKit / Widget Builder<br/>(distribution channels)"]
        O1 --> O2 --> O3 --> O4
    end

5. Why the harness layer specifically is becoming vendor-locked

5.1 The "enterprises will pay for this eventually" learned behavior

flowchart LR
    GH["GitHub<br/>(Enterprise)"] -->|"ACL/SAML/Audit"| ENT1["Enterprise ISMS requirement"]
    Slack["Slack<br/>(Enterprise Grid)"] -->|"DLP/eDiscovery"| ENT1
    Notion["Notion<br/>(Enterprise)"] -->|"SCIM/audit logs"| ENT1
    Claude["Claude Team/Enterprise"] -->|"RBAC/SSO"| ENT1
    OpenAI["ChatGPT Business/Enterprise"] -->|"audit logs/usage analytics"| ENT1

• Free / MAX = personal coding assistant

• Team / Enterprise = harness governance (audit, tracing, multi-seat, policy gates)

5.2 OpenAI's "harness = part of the model's training distribution" claim

6. Infra costs are exploding — both cloud and local get more expensive

6.1 LLMs are getting cheaper like air, but "everything around them" is getting more expensive

"AI's enemy is time. Today's Pro-tier inference becomes free-tier in 6 months."

— Memorizer: ChatGPT Pro 90% Discount Promotion in South Korea (대한민국)

• GPU: for LLM inference

• Memory / SSD / NAND: local PC prices spiking (as users have noted)

• Compute (BYOC): OpenAI explicitly says bring your own compute for the long-horizon harness

6.2 Google's counter — TurboQuant + the Jevons Paradox

"If inference truly gets cheaper, the freed memory gets used for applications that were previously priced out — total memory demand could explode rather than shrink."

flowchart LR
    A["Efficiency 6×↑<br/>(TurboQuant)"] --> B["Unit cost ↓"]
    B --> C["Agent workloads<br/>explode"]
    C --> D["Total memory · GPU · power<br/>demand ↑↑"]
    D --> E["Local PC parts<br/>+ cloud bills<br/>both rise"]

6.3 Cloud vs local — both get more expensive at the same time

7. So what should we prepare?

7.1 Three scenarios

flowchart TD
    Choice["Our choice"]
    Choice --> S1["Scenario A<br/>Play on top of vendor-locked harness"]
    Choice --> S2["Scenario B<br/>Design a native harness ourselves"]
    Choice --> S3["Scenario C<br/>Hybrid — partly on-device, partly frontier model"]

    S1 -->|"Pros"| S1P["Fast start, full-stack tooling provided"]
    S1 -->|"Cons"| S1C["Hostage to license & pricing tier, model lock-in"]

    S2 -->|"Pros"| S2P["Full control, multi-vendor possible"]
    S2 -->|"Cons"| S2C["Engineering cost, learning curve"]

    S3 -->|"Pros"| S3P["Cost-optimal, risk diversified"]
    S3 -->|"Cons"| S3C["Vendor harness may not allow this kind of bridging"]

7.2 The trap — "forced upgrade to Team/Enterprise"

7.3 Recommended internal actions

8. Closing — Which side are we betting on?

"Every component in a harness encodes an assumption about what the model can't do on its own."

1. "Should we become native harness engineers?" — Yes. But not at every layer — only at the domain-specific layer (e.g., our domain's Evaluator, our 5-axis evaluation, our governance policies).

2. "Is it more rational to just play on top of the vendor harness?" — Partly yes. But keep three escape hatches reserved in advance: MCP, harness tracing, and data governance.

References

External

• Anthropic — Designing harnesses for long-running applications, Prithvi Rajasekaran, 2026.3.24
<https://www.anthropic.com/engineering/harness-design-long-running-apps>

• OpenAI — The next evolution of the Agents SDK, 2026.4.15
<https://openai.com/index/the-next-evolution-of-the-agents-sdk/>

• OpenAI — Harness engineering (KO)
<https://openai.com/ko-KR/index/harness-engineering/>

• TechCrunch — OpenAI updates its Agents SDK to help enterprises build safer, more capable agents, 2026.4.15
<https://techcrunch.com/2026/04/15/openai-updates-its-agents-sdk-to-help-enterprises-build-safer-more-capable-agents/>

• Daum News — TurboQuant and the Jevons Paradox, 2026.3.29
<https://v.daum.net/v/20260329050205578>

Internal (Memorizer)

• Claude Code — Part 1: Harness, AI Agent, Agentic Loop — 악분, 2026.3.22

• HARNESS — Agent Team & Skill Architect Architecture — Minho Hwang, 2026.3.18

• Harness-driven dev journey v0.0.1–v0.0.5 — Project Generation, Validation, Improvement Cycle

• OpenAI DevDay 2025: ChatGPT's Super-App Strategy

• OpenAI ChatGPT Pro 90% Discount Promotion in South Korea (대한민국) (KakaoTalk)

• Anthropic Academy & Ambassadors — Part 2: Pros, Cons, and OpenAI Comparison

• For CEOs Considering AX (AI Transformation) — ChannelTalk CAIO Lee Kyung-hoon

• Software Development in the AI Era: The Return of Architecture

TECH LINKS

• 𝕏 @webnori

• 📘 Akka Labs (Facebook group)

• 📝 More writings — webnori wiki