Capital & Compute
· ai agents· harness engineering· build vs buy· claude code

Build Your Own Agent Harness or Buy Claude Code?

Should you build your own AI agent harness or adopt Claude Code and Codex? A cost and control decision guide: adopt, extend, or build.

By Capital & Compute

Most teams asking “should we build our own agent harness” are answering the wrong question. The real choice in 2026 is not build-versus-buy. It is a three-way split, and the middle option is the one almost everyone actually wants.

Here is the decision rule, before any of the reasoning behind it:

  1. Adopt as-is. Use Claude Code or Codex out of the box. This is right for most individuals and most teams shipping general software.
  2. Adopt and extend. Build on top of a harness through its SDK, plus the Model Context Protocol, hooks, subagents, and skills. This is the default the moment you need proprietary tools, a domain workflow, or organization-wide guardrails. It is what “we need something custom” almost always means.
  3. Build from scratch. Write your own loop, context management, tool layer, verification, and guardrails. This is correct only under a few narrow conditions, and cost is never one of them.

The reason the middle option gets skipped is that “build” and “buy” sound like the only two doors. They are not. Modern harnesses now expose so much of their own machinery that you can get deep customization without rebuilding the engine. This guide walks the three paths on cost, control, and lock-in, and shows why the build-from-scratch case has narrowed to a short list.

+13.7 pts
Benchmark gain from the harness alone
Fixed model; only the scaffold changed (LangChain)
$150–250
Monthly cost per developer to adopt
Anthropic's stated Claude Code range
~$120k+
Modeled cost to build a harness from scratch
About six engineer-months before any product work

What a harness is, and why it is the expensive part

An agent harness is everything wrapped around the language model that turns a single prompt into an agent that does work: the loop that lets it act and observe, the context it reads, the tools it can call, the checks that grade its output, and the limits on what it may touch. The model supplies raw capability. The harness decides whether that capability lands, and it sets the bill. The companion piece on how to scaffold a harness layer by layer breaks down each of those five layers and the token cost each one controls.

The reason this matters for a build-versus-buy decision is that the harness is where the engineering difficulty lives. In late 2026 a team at LangChain took a coding agent scoring 52.8% on Terminal Bench 2.0, a placement outside the top 30, and pushed it to 66.5% and a top-five finish without switching models, documented in LangChain’s own write-up. The full argument, and why the harness now outweighs the model as a differentiator, is laid out in the piece on Hermes Agent and why the harness beats the model.

Rebuilding a harness from scratch means rebuilding the part that took the frontier labs years to get right.
The core cost of the build-from-scratch path

That 13.7-point swing is the whole case against a casual build decision. The scaffold was worth more than a generation of model progress. When you choose to build from scratch, that is the work you are signing up to recreate: the loop control, the context compaction, the tool orchestration, and the verification that produce results like it. Frontier vendors have poured years of engineering into exactly this. The practitioner Addy Osmani compressed the lesson in his 2026 essay Agent Harness Engineering: “a decent model with a great harness beats a great model with a bad harness.” A from-scratch harness starts as the bad harness.

The middle path most teams miss: adopt and extend

The false binary hides the option that fits the most common real requirement. When a team says it needs a custom agent, it usually means one of three concrete things: connect the agent to internal systems, teach it a specific workflow, or enforce organization rules on what it can do. None of those requires building a loop from scratch. All three are now first-class extension points in the tools you would otherwise be competing against.

The Claude Agent SDK gives you “the same tools, agent loop, and context management that power Claude Code,” callable from Python or TypeScript, per Anthropic’s documentation. On top of that loop it exposes the extension surface directly: custom tools and external systems through the Model Context Protocol, lifecycle hooks (PreToolUse, PostToolUse, SessionStart, and more) to validate or block behavior, subagents with their own context windows for delegated work, a permission model, and resumable sessions. You supply your tools and rules; the hard runtime is already built and maintained by the vendor.

Codex offers a parallel path. Per OpenAI’s documentation, running codex mcp-server exposes Codex to any MCP client through two tools, codex() to start a session and codex-reply() to continue one, and OpenAI frames the pattern as a way to build “deterministic, reviewable workflows” on Codex “rather than building orchestration from scratch.” Codex also adds skills, a hooks system, and a Python SDK for embedding it in CI or custom tooling. The extension primitives on the Claude Code side, and how skills, hooks, subagents, and MCP fit together, are covered in depth in the Claude Code harness guide, with more advanced patterns in harness engineering techniques beyond MCP.

The three paths on the dimensions that decide it

Laying the three approaches against the axes that actually drive the decision makes the pattern visible. Adopt and extend wins or ties on almost everything a buyer cares about. Build from scratch wins on exactly two things, and pays for them everywhere else.

Adopt, extend, or build: a decision matrixA six-row matrix rating three approaches. Time to first value: adopt minutes (advantage), extend days (advantage), build months (drawback). Upfront cost: adopt none (advantage), extend low (trade-off), build high (drawback). Ongoing maintenance: adopt on the vendor (advantage), extend shared (trade-off), build all yours (drawback). Customization: adopt limited (trade-off), extend high (advantage), build total (advantage). Vendor lock-in: adopt high (drawback), extend moderate (trade-off), build none (advantage). Harness quality: adopt frontier-grade (advantage), extend frontier plus your tools (advantage), build you rebuild it (drawback).AdvantageTrade-offDrawbackDimensionAdopt as-isAdopt and extendBuild from scratchTime to first valueMinutesDaysMonthsUpfront costNoneLowHighOngoing maintenanceOn vendorSharedAll yoursCustomizationLimitedHighTotalVendor lock-inHighModerateNoneHarness qualityFrontierFrontier +Rebuild it
Adopt, extend, or build: a decision matrix
DimensionAdopt as-isAdopt and extendBuild from scratch
Time to first valueMinutesDaysMonths
Upfront costNoneLowHigh
Ongoing maintenanceOn vendorSharedAll yours
CustomizationLimitedHighTotal
Vendor lock-inHighModerateNone
Harness qualityFrontierFrontier +Rebuild it
The three approaches scored on the six dimensions that decide the call. Green is an advantage for the buyer, amber a trade-off, red a drawback. Adopt-and-extend carries no red; build-from-scratch buys control and independence at the cost of everything else.Source: Capital & Compute analysis, from the Claude Agent SDK and Codex documentation and Anthropic's Claude Code cost guidance

Two cells deserve a note. Under vendor lock-in, adopting as-is scores worst because your workflow lives inside one vendor’s product; extending on an SDK is only moderately better, since Anthropic’s terms bar third-party products built on the Agent SDK from offering Claude login and require API-key billing, per its documentation. Building from scratch is the only genuinely model-agnostic path, which is a real advantage if independence is a hard requirement. Under harness quality, building scores worst not because your engineers are weak but because you begin from zero against a moving frontier target.

The economics: building does not save money

The most common unspoken reason teams reach for “build our own” is a belief that owning the harness will be cheaper than paying a vendor forever. Run the numbers and the belief inverts. Building is the most expensive path on almost any realistic horizon, for a structural reason: building the harness does not reduce the model bill. You still rent the model by the token whether you wrote the loop or the vendor did. Anthropic’s own Claude Code cost guidance puts the average developer between $150 and $250 a month in model usage, and that line item is roughly the same under all three approaches. What changes is the fixed cost you stack on top.

The chart below models a five-developer team over two years. All three paths carry the same modeled model bill, about $1,000 a month for the team. Adopting as-is adds nothing else. Extending adds a modest one-time build and light maintenance. Building from scratch adds a large upfront cost and a standing maintenance burden, because a custom harness has to be kept current as models, APIs, and pricing shift underneath it. The assumptions are stated in full below the figure; the point is the shape, not the decimal.

Cumulative cost over two years: adopt, extend, or buildCumulative cost by month for a five-developer team. Adopt as-is starts at zero and rises to about 24,000 dollars over 24 months on model usage alone. Adopt and extend starts near 15,000 dollars and reaches about 51,000 dollars. Build from scratch starts near 120,000 dollars and reaches about 336,000 dollars. The build line stays highest throughout and never crosses below the others.$0$100k$200k$300k$400kStartMo 6Mo 12Mo 18Mo 24Build from scratch$336kAdopt and extend$51kAdopt as-is$24k
Cumulative cost over two years: adopt, extend, or build. Cumulative cost by month for each approach.
MonthAdopt as-isAdopt and extendBuild from scratch
0$0$15k$120k
6$6k$24k$174k
12$12k$33k$228k
18$18k$42k$282k
24$24k$51k$336k
Modeled cumulative cost for a five-developer team. All three paths pay the same model usage; building adds a large fixed cost and standing maintenance that adopting never incurs. Illustrative, from the stated assumptions, not a quote.Source: Capital & Compute cost model, grounded in Anthropic's stated Claude Code usage range and stated engineering-cost assumptions

The build line never dips below the others. There is no payback month, because there is no saving to pay back: the model cost is a wash across all three, so the extra engineering is pure added cost. That is the honest financial verdict. If you build, you are spending more to get control, not spending more to eventually save. For a fuller treatment of the recurring model bill that sits under every path, see what it costs to run an AI agent in production and the per-task breakdown in Claude Code pricing, cost per task.

When building from scratch is the right call

Building is sometimes correct. The trigger is never cost, and rarely customization, because the SDK path covers most customization. It is one of a short list of structural reasons:

  • A domain the coding harnesses cannot model. Claude Code and Codex are built for software work: files, a shell, a repository. If the agent operates in a domain with a fundamentally different loop, action space, or state, for example robotics, a trading system, or a physical-process controller, the coding harness is the wrong shape and extending it fights the grain.
  • Hard compliance or air-gap requirements the vendor cannot meet. If code and context legally cannot leave your environment, or a regulator requires full control of the runtime, and no vendor deployment option satisfies it, you may have to own the whole stack. Note that adopting still has room here: the Agent SDK runs the loop inside your own process and supports Bedrock, Vertex, and Azure deployments, per its documentation, so confirm the vendor cannot meet the requirement before building.
  • The harness is your product. If you are selling the agent itself, its loop and orchestration are your differentiator and your moat, not a cost center, so building is the business.
  • You are researching the loop. If the object of the work is the agent architecture, the harness is the experiment.

Absent one of those, extending a frontier harness beats building one, and the cost model is why. This site’s head-to-head on Claude Code versus Codex on cost and capability is the right next read for choosing which harness to adopt or extend, and the AI coding agents hub maps the full field with the current coding-plan pricing. For the sourcing standards behind the figures here, see the editorial standards.

Frequently asked questions

Frequently asked questions

Is it cheaper to build my own agent harness than to pay for Claude Code?
No. Building does not reduce the model bill, since you still rent the model by the token, and it adds a large upfront engineering cost plus ongoing maintenance as models change. On a realistic multi-year horizon the build path is the most expensive of the three, so build for control or compliance reasons, not to save money.
What is the difference between adopting and extending a harness?
Adopting means using Claude Code or Codex as shipped. Extending means building on top of the same harness through its SDK, the Model Context Protocol, hooks, subagents, and skills, so you add proprietary tools, workflows, and guardrails while the vendor still maintains the loop, context management, and verification. Extending gives most of the customization of building with a small fraction of the cost and risk.
Can I use my own models and tools without building a harness from scratch?
Mostly yes for tools, partly for models. The Claude Agent SDK and Codex both let you add custom tools through MCP and control behavior with hooks and subagents. Model choice is more constrained: each vendor SDK is built around its own models, so full model-agnosticism is one of the few genuine reasons to build your own harness.
When does building your own agent harness actually make sense?
When the agent works in a domain the coding harnesses cannot model, when hard compliance or air-gap rules leave no vendor option, when the harness itself is the product you sell, or when you are researching the agent loop. Outside those cases, extending an existing harness is the better call.

Sources

Anthropic. Agent SDK overview. Claude Code documentation (vendor documentation), verified 2026-07-13. https://code.claude.com/docs/en/agent-sdk/overview

Anthropic. Manage costs effectively. Claude Code documentation (vendor documentation), verified 2026-07-13. https://code.claude.com/docs/en/costs

OpenAI. Running Codex as an MCP server with the Agents SDK. Codex documentation (vendor documentation), verified 2026-07-13. https://learn.chatgpt.com/docs/mcp-server

LangChain (2026). Improving deep agents with harness engineering. LangChain blog (vendor/practitioner blog), verified 2026-07-05. https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering

Osmani, A. (2026). Agent Harness Engineering. Practitioner essay, verified 2026-07-13. https://addyosmani.com/blog/agent-harness-engineering/

Subscribe to Capital & Compute

Source-backed analysis of what AI compute really costs, sent when a new post goes live.

No spam. Unsubscribe anytime.

← Back to Coding agents