What Happens When You Change Tool Calling from "JSON" to "Code"? Empirical Results Reveal the Build-or-Buy Inflection Point

---

TL;DR

• Empirical testing across 14 models on BFCL v4: simply switching tool calling from “calling via JSON” to “calling as code” (Programmatic Tool Calling) let 11 of 14 models match or exceed baseline performance, with clear gains in parallel execution and long-context retention.
• Japan’s Ministry of Justice has explicitly designated “voice” as a personality right protected alongside likeness and name rights, clarifying the legal risk contours for voice-cloning products.
• Preferred Networks’ domestic LLM “PLaMo 3.0 Prime” is now available on Sakura Internet’s AI inference platform, but pricing is undisclosed and application-based, making it impossible to evaluate for now.

Headline: What Happens When You Change Tool Calling from “JSON” to “Code”? Empirical Results Reveal the Build-or-Buy Inflection Point

Technical take

The study compared, across 14 LLMs on the established benchmark BFCL v4, the conventional approach of calling tools in JSON format (native JSON tool calling) against an approach where tools are exposed as typed Python stubs and the model calls them as code (Programmatic Tool Calling, PTC) (arXiv). With PTC, chaining and parallelizing multiple tool calls can be written naturally within code, allowing execution and result retrieval to be consolidated into a single agent turn. As a result, PTC matched or exceeded the JSON approach in 11 of 14 models, with the GPT-5.6 family showing a +10.6% improvement over the JSON baseline. Under parallel fan-out conditions calling multiple tools simultaneously, PTC matched or exceeded JSON in 13 of 14 models, and even under “context rot” conditions—where accuracy degrades as context grows longer—PTC remained stable while the JSON approach degraded by an average of 2.3%. However, PTC did not outperform in 3 of the 14 models, and the paper also notes that “performance gains track the capability of the model generation.” As this is based on a single benchmark and an unreviewed arXiv preprint, the findings should be read with appropriate caution.

Business take

What this result points to is a build-or-buy inflection point: whether to build your own tool-calling orchestration layer (JSON calls plus retry plus parallel control) in-house, or to rely on the standard features offered by the model/platform side. If PTC-equivalent calling methods get absorbed into platforms as a default feature going forward, the JSON-based parallelization and retry mechanisms being custom-built in-house today risk becoming reinvented wheels.

Implications and positioning

If you’ve already built out the scaffolding for JSON calls plus retry plus parallel control in-house, no further investment is needed there. If you have processes that chain multiple tools together, start by porting to a model/SDK that supports code-execution-based calling, and verify whether success rates during parallel execution actually improve in your own workflow. There is essentially no impact on products that only make single, standalone tool calls.

Other Key Topics

1. Unauthorized voice use: Ministry of Justice states for the first time it “can constitute a publicity rights violation”

On August 7, Japan’s Ministry of Justice released interpretive guidance stating that voice actors’ “voices,” like their likeness and name, are subject to protection under personality rights (ITmedia). The guidance positions voice as “personally identifying information and a symbol of an individual’s personality,” making it subject to protection both under “publicity rights”—the exclusive right to commercially exploit a celebrity’s customer-drawing power—and under the personality-rights-derived “right not to be used without authorization.” As examples, it cited generating and monetizing videos of a character singing in the voice of the voice actor who plays them using generative AI, and publishing audio in which a voice resembling a voice actor’s reads obscene text. For businesses, the guidance states that offering a paid generative AI service that markets, as a “selling point,” the ability to reproduce a specific voice actor’s voice could also be judged as a violation of publicity rights. The backdrop includes voice actor Kenjiro Tsuda’s lawsuit against TikTok’s operating company over a video that imitated his voice without authorization.

The technology itself (voice cloning) remains unchanged as existing technology, but what’s new from a business standpoint is that data governance design—whose voice is used, and how consent is obtained—has now become a clearly delineated subject of legal risk. Business models that market themselves on “voices resembling famous voice actors” risk having the business itself named as the infringing party.

Implications and positioning: If you have, or are considering, a voice synthesis or voice cloning product, you should stop demos and marketing that imitate celebrity voices immediately. Conversely, for businesses that hold consent-obtained voice databases, this guidance becomes a compliance-side moat.

2. Domestic LLM “PLaMo 3.0 Prime” arrives on Sakura Internet’s AI inference platform, but pricing is unknown until you apply

On August 4, Sakura Internet announced that it has begun offering Preferred Networks’ LLM “PLaMo 3.0 Prime” on its generative AI inference API platform “Sakura no AI Engine” (ITmedia). Access requires an application and approval from the provider, and the service is not available under the free plan. Pricing is displayed only to approved users. PLaMo 3.0 Prime is a reasoning model built on a pretrained model developed through joint research with NICT, extending context length from 64k to 256k tokens and using a proprietary tokenizer aimed at improving token efficiency and cost performance. In PFN’s internal evaluation, the model reportedly delivered competitive results in Japanese instruction-following, coding, and tool use compared to open models in the same performance tier such as Qwen3.6-27B and gpt-oss-120b, and closed models in the same price tier such as GPT-5.4 mini and Claude Haiku 4.5.

This performance comparison is PFN’s own internal evaluation, and no third-party verification is noted. What’s business-relevant here is less the performance claims than the distribution model: excluding a free plan and adopting an application-based, price-undisclosed system suggests a strategy of negotiating individual pricing with enterprise customers.

Implications and positioning: Keep this on your radar as a candidate for customers (government agencies, financial institutions, etc.) that require Japanese data sovereignty or compliance with domestic regulations. Since cost comparison is impossible while pricing remains undisclosed, there is no reason to switch now.

3. Comparing LLM agent strength: a statistical method that makes poker evaluation 74x cheaper

Determining which of two AI agents is stronger requires playing enough games to overcome the element of luck, and each additional game incurs costs, inference expense, and expert time. This research proposes AV-AIVAT, which combines AIVAT—a variance-reduction technique for imperfect-information games—with a statistical method that permits stopping once sufficient evidence has accumulated (Anytime-Valid Confidence Sequences) (arXiv). Across data from 71,439 paired hands of Heads-Up No-Limit Hold’em (HUNL) played among 15 LLM agent configurations, AIVAT achieved a median 54x reduction in variance. At a 95% confidence level and a target precision of ±1BB, uncorrected raw data reportedly required a median of 74 times as many hands as AIVAT-corrected data before statistically valid stopping became possible.

Technically, this amounts to one solution to the evaluation-cost problem in AI agent development, in the sense that it cuts evaluation costs while rigorously guaranteeing a statistically valid “stopping point.” However, the demonstration is limited to poker, an imperfect-information game, and its transferability to general task evaluation remains unverified.

Implications and positioning: If evaluation costs are piling up in agent-vs-agent A/B testing, the design philosophy of “stopping once sufficient evidence has accumulated” is worth taking away. However, transfer to tasks beyond poker remains unverified, so there is no reason to adopt this immediately.

4. 97% of job-seeking students use generative AI; 20% struggle when interviewers probe deeper

In a survey of university students set to graduate in spring 2027, only 3% said they do not use generative AI in job hunting (ITmedia). This “non-use” rate has fallen over time: 55% for spring 2025 graduates, 24% for spring 2026 graduates, and 3% for spring 2027 graduates. The most common use case was “proofreading/improving entry sheets” (over 60%), followed by “self-analysis” (roughly 50%). When pressed by interviewers to elaborate on details written in their entry sheets, a combined 20% of students said they could not clearly answer “about half” or more of such follow-up questions. The survey was conducted in June 2026, drew 346 responses, and was published by recruiting firm Port.

This data indicates that entry sheets—the input data of the hiring process—are becoming homogenized as AI-edited text, suggesting that companies’ selection processes need to shift their evaluation axis away from assuming “text the candidate wrote themselves” and toward improvised, in-depth follow-up questioning during interviews.

Implications and positioning: If you’re considering an HR/recruiting product, it makes more sense to bet on “redesigning the evaluation process around the assumption of AI assistance” (generating follow-up questions, verifying consistency in interviews, etc.) rather than on “AI-use detection.” A standalone AI-use detection feature is becoming nearly meaningless given the current 97% usage rate. That said, this survey has n=346 and comes from a specific research firm, so its representativeness of the broader population should be read with appropriate caution.

Worth Trying This Week / Hype Safe to Ignore

Worth trying this week: If you have workflows involving chained tool calls, switch a portion of them to a PTC-supporting SDK and measure the actual difference in parallel-execution success rates versus the JSON approach.

Hype safe to ignore: PLaMo 3.0 Prime’s claim of matching or exceeding overseas models cannot be evaluated while pricing remains undisclosed, and is not a reason to switch now. The 74x figure from AV-AIVAT is a demonstration in the specialized domain of poker, and cannot be assumed to directly apply to general-purpose agent evaluation.