How a Claude-powered agentic harness helped safely engineer a market-ready product for a highly-regulated sector.

platform3 (opens in new tab), an AI-native accounting startup, is rethinking statutory financial statements preparating. The company is enabling businesses to use AI to bring control, and massive speed and efficiency gains, to processes mired in legacy technology.

With the deadline for UK companies to file accounts using commercial software in iXBRL format rapidly approaching, platform3 had already produced a proof-of-concept that enables businesses to leverage AI to review the statutory accounts against the Financial Reporting Standards with citations.

But to get this into the workflow of accountancy firms, platform3 needed production-grade engineering—tested, reviewable, defensible code—that competitors could not simply copy, delivered rapidly and within a startup’s budget.

To achieve platform3’s ambitious targets, Softwire augmented a small team of senior engineers with AI agents working inside an agentic harness built using Anthropic’s Claude. 

This harness made the code produced by the AI agents more trustworthy by creating the environment (the context, tools, quality sensors, workflow and improvement loop) that the agents work in, thus setting the guardrails that allowed agents to work safely, at speed. 

The agentic harness operated under the following principles: 

Design before code: before an agent writes a line, a human reviews a sub-minute design package. Saying no costs nothing. 

Agents do everything except merge: an agent plans, implements, tests and reviews its work and raises the merge request. Merging is human-only, always, enforced by tooling. 

Correctness is computable: in a regulated domain, the checks that matter are deterministic—journals balance, money is never a float, depreciation reconciles. They live in code, not prompts. 

Cite, don’t recall: agents may not state an accounting rule from memory. They search the rulebook and cite the paragraph. 

More review, not less: a small team meant strengthening the process, not relaxing it. 

Weekly retrospectives: regular retros tuned the harness; ticket sizing, convergence speed, merge conflicts and front-end quality improved week-on-week.

Given the highly-regulated domain platform3 operates in, the AI inside their product needed to be 'harnessed’ with a layer of context, guardrails and gates to ensure that it was constrained to a strict output contract that obeyed tight statutory rules.  

This runtime harness was permitted to propose findings and extractions, but never do arithmetic, set a severity or produce a figure that posts. Everything it claims is then checked by code that cannot hallucinate. Totals are cast, citations located in the source document, and figures computed deterministically from source data, ensuring trusted output that can be put in front of an auditor. 

To deliver confidence in both the design process and the product itself, the work was reviewed by both AI and humans at multiple stages. The harness stipulated human review before any code was written. Every change was then checked by AI review, automated tests, and a human reviewer. Finally, only a human can approve merging the code into the product. As the team’s pace increased, the reviews became tighter. 

The harness was deliberately split into three parts, to ensure that the AI can never weaken its own controls: 

The checker: a deterministic quality-gate runner over every change set. No AI in it, by design. 

The packs: reusable rule presets the checker consumes: accounting, AI extraction, .NET, Angular and Python. 

The driver: the container, workflow skills and AI review, powered by Anthropic’s Claude. The only place AI lives. 

And of course, as well as the harness, the team delivered the product itself: template-free, taxonomy-driven document generation. This included a transparent AI review that shows which check passed or failed – a vital feature singled out by platform3 users. 

Seven weeks after this project kicked off at a half-day hackathon, platform3 were selling a production model, delivered under-estimate and in a fraction of the time it would have taken to build a non-AI automated product.  

The company keep their own version of the harness, tuned to IFRS, FRS101 and FRS 102, for continued development. The resulting product also contained a number of thoughtful AI use cases to streamline accountancy workflows, including contract extraction, AI review, and intelligent roll forward from the previous year. The user is even able to configure which models and model providers they use per use case between various Anthropic and OpenAI models to balance speed, cost and performance.