AI in SDLC: Model doesn’t matter

Did you know that Claude Code can deliver up to 46% better results just by changing how you run it? In almost every conversation about AI in software delivery, sooner or later, someone asks, “Which model are you using?” You may wonder why we consider this a less interesting question in the whole equation.

Let’s have a brief look at the flow of this post: at first, a little recap of what the benchmarks actually say, so we are on the same page, then the results we achieved on ourselves as client zero, followed by the methodology we use to repeat them at your place, and we are going to finish with what all of this means for your business.

Harness takes the stage

The optimization mentioned can be done more effectively than just switching the model, and it relates to a harness.

In the context of coding agents, a harness is the layer around the model that determines how the model operates:

• what context it receives,
• which tools it can use,
• how it interacts with the codebase,
• how it executes and verifies changes,
• and how the overall task is orchestrated.

A 2026 survey of agent systems (arXiv:2606.20683) compares the same Claude models across different harnesses on a well-established benchmark (SWE-bench Verified). The numbers jump from 33.6% to 49.0% of tasks solved, approximately a 46% relative gain when only the scaffolding around the model changes. Even at the frontier, harness choice still moves the score from 76.8% to 80.9%.

And this idea is starting to show up beyond software engineering benchmarks. In August 2026, NVIDIA demonstrated just how much the surrounding system can matter on ARC-AGI-3: the same Claude Opus 5 reportedly went from 30% to 100% after being placed inside a more sophisticated harness, including better memory management and a supervisory component.

As TechCrunch put it,

The harness – rather than the model itself – is increasingly becoming the real hero of long-horizon agentic tasks.

TechCrunch, 21.08.2026

What we’ve learned in production is now confirmed by the research community. The model – this legendary subject of the very first “what do you develop with” questions is a commodity everyone can rent at the same prices, through the same APIs.

Yet with the same model, we’ve seen results ranging from statistical noise to a tenfold acceleration. Having investigated, we found the harness to be far more important, and we treat it as the engineering environment you build around your project so that it works within your SDLC, not beside it.

The evidence keeps piling up

In our recent analysis of the two newest flagship models, Beyond the flagship: what Fable 5.1 and GPT-6 Astra tell us about AI-native delivery, we went through everything measurable published in their first week.

The most durable findings were not about either model, but about the systems around them:

• Same weights, different scaffolding. ARC Prize reported GPT-6 Astra at 62.7% on ARC-AGI-3 under a provider-neutral harness and 98.6% with OpenAI’s own Provider Adapter. The difference is that the adapter keeps the model’s reasoning state between actions and manages the context itself.
• Benchmarks increasingly measure systems. Real-SWE from Specific Labs, built on private production repositories, scores model-and-harness pairs rather than models: Fable 5.1 in Claude Code leads at 38.8%, and Astra in Codex CLI reaches 33.8%, with no configuration above 40%. The most common reason for failure is a missed requirement, exactly the kind of gap that project-specific context and verification steps are meant to close.

This is the nuance behind our title. It is not enough to pick a harness, just as it is not enough to pick a model. What matters is how well the harness configuration fits your codebase, your workload, and your delivery process, and whether you measure it on your own tasks rather than on a public leaderboard.

We were our own Client Zero

Such research is time-consuming, which is why Sii Poland comes with a good part of it done for you already. We simply tested it on ourselves.

Our internal delivery organization (more than 115 applications and platforms and over 280 integration modules across .NET, Angular, Microsoft Dynamics 365, WordPress, and heavyweight integration middleware) became Client Zero.

Every AI-augmented team has access to the same models. What we engineered was the harness:

• organizational developer kits with project-specific context rules,
• specialized agents for implementation, review, testing, and bug reporting,
• MCP servers wired into real infrastructure,
• and safeguards that block secrets and require developer review.

The results are public already and can be found here: Faster software delivery with Claude Code – Sii Poland case study.

Anthropic’s methodology brought to your place

You may wonder how transferable all of this is. Pretty well, as it turns out, because we follow the deployment methodology Anthropic itself publishes for enterprises.

1. Let’s wire the Foundation on your platform of choice, the enterprise controls (SSO, etc.), directory sync, audit logs, data retention, and only then build the first project harness. This is the phase where most organizations stall for months. Still, we narrow it down to days, supporting the process with our SDLC Bootstrap framework that compresses the harness configuration stage into a guided, reviewable run that mines your repository and connects the dots between the code and your actual delivery process.
2. Pilot puts the harness in the hands of champions on real backlog items for a few weeks, with weekly check-ins and two hard metrics: cycle time saved, and the share of AI output accepted without substantial revision. Scale turns what worked into organizational assets, namely skills, plugins, and standards rolled out team by team. This way, every subsequent project starts from accumulated knowledge rather than from zero.

What the engagement covers

Basically, the work is broader than installing a tool, because the harness is broader than a tool.

• Security-wise, the managed permission policies that local configuration cannot override, hooks that stop secrets from ever reaching the repository, and automated security review on every pull request.
• For the sake of Governance with all the SSO and SCIM, audit logs, usage and cost monitoring, model-version pinning, and a non-negotiable human-acceptance rule, it is a developer who approves every change.
• Through the means of MCP integrations, we provide controlled, auditable connections between Claude and the systems your delivery actually runs on – Jira or Azure DevOps, databases, logs, and observability, internal component libraries – checked into the repository so the whole team benefits and so benefits the control and auditability of these connections.
• We endow you with the standards, skills, and training that enable the tenth team’s rollout faster than the first.

Artificial Intelligence

We deliver AI solutions tailored to your business, driving efficiency and boosting productivity within your teams.

AI offering

On your infrastructure – AWS, Azure, or GCP

None of this requires your code or prompts to leave your cloud, which, for many of our clients, is the deciding factor. Claude Code deploys natively against Amazon Bedrock on AWS, Google Cloud’s Agent Platform (formerly Vertex AI) on GCP, and Microsoft Foundry on Azure.

The best of two worlds: frontier models and AI services when you want them, your own infrastructure always. It’s all billed through your existing cloud account, authenticated with your IAM or Entra ID, and audited by CloudTrail, Cloud Audit Logs, or Azure Monitor, thereby strengthening the processes and FinOps running in the background, with all corporate proxies and LLM gateways supported where network policy requires them.

Our delivered solutions are sized to fit, and so is the mode of collaboration

There are three easy ways to start.

1. A discovery workshop – e.g., half a day in which we map one project’s SDLC and show you exactly what its harness would look like.
2. A pilot – two to four weeks in which we stand up the Foundation on your cloud and put a harness in front of one real team, on a real backlog, with the metrics to prove it.
3. A full program – which is essentially Foundation to Scale across your delivery organization.

Summary

All right, let’s have a quick summary. The benchmarks say the harness moves the needle more than the model does. We proved it to ourselves and published the numbers. We repeat it with Anthropic’s methodology, on your cloud, with security and governance built in rather than bolted on. The Client Zero was us.

You can join other clients and us – reach out and tell us how you want to work.