AI Agents, Engineering Jobs and the Future of Software: Augment, NEAR and Modular

Matthew Berman and Nick Wentz open the March 6, 2026 edition of The Future Live with AI news, then welcome guests from Augment Code, NEAR Protocol and Modular. Their conversations focus on coding agents, personal agents, software infrastructure and the uncertain effect of AI on engineering jobs.

Wentz opens with OpenAI's GPT-5.4 launch in ChatGPT, the API and Codex. Berman describes it as bringing general knowledge and reasoning together with coding capabilities previously associated with separate models. He says his initial tests are positive, while noting that he is still working out its strengths and weaknesses.

The hosts discuss whether one model for coding, tool use and general work makes selection simpler for enterprises and knowledge workers. Wentz cites OpenAI's reported reductions in factual errors against GPT-5.2: 33% at the level of individual claims and 18% at the response level. Those figures concern OpenAI's evaluated prompts, not a guarantee of accuracy on every task.

Berman says hallucinations remain a practical concern and users still need to check results. Asked whether the release timing was intended to change the conversation after negative news about OpenAI, he says any answer would be speculation; a model release is prepared over a long period.

The hosts next discuss Apple's lower-profile approach to AI infrastructure spending. Berman proposes that improving chips and smaller models could support a hybrid future, with some tasks on personal devices and the hardest work in large cloud models. He presents this as a developing idea rather than a settled forecast.

Wentz argues that Apple's existing devices, including headphones, could become valuable interfaces for voice AI. Berman offers the counterargument: if demand for powerful cloud inference grows faster than local devices improve, Apple could remain dependent on outside partners. The hosts also discuss whether acquisitions might help, without claiming that any particular deal is planned.

The first guest is Vinay Perneti, Augment Code's VP of engineering. He says Augment began with the problem of giving AI systems relevant codebase context and now builds agents for professional engineering teams. His vision is small teams of engineers coordinating larger teams of agents.

Perneti describes Augment's beta product Intent as an attempt to organize work around a desired outcome and a team of agents rather than around a conventional IDE. Berman asks what happens when developers move further away from reading code line by line; Perneti answers that architecture, design principles and verification become more important.

Berman challenges the analogy with a compiler: unlike a deterministic compilation step, the same instruction to different models can produce different code. Perneti agrees that verification is critical and says Intent includes a verification agent designed to compare the implementation with the developer's intent. He describes tighter verification as continuing research, not a solved problem.

Asked how Augment verifies work in practice, Perneti describes unit and system tests, CI/CD, canary releases and feature flags for rollback. His team also uses a hierarchy of specifications: people agree on the system design and component-level requirements, then use agents to refine implementation specs and check the resulting work before the pipeline runs.

Wentz asks how an organization moves to an agent-first workflow without ignoring the human impact. Perneti says that after the winter break he noticed experienced engineers increasingly delegating code to agents and working in parts of the codebase they might previously have avoided. He treats this as evidence of a shift in practice at his company, not a measurement of all engineering teams.

Perneti says agents remain imperfect but predicts much longer tasks will become feasible as capabilities improve. He recounts an Anthropic researcher's forecast of tenfold longer tasks by the end of 2026; that is a forecast, not an established result. This possibility raised uncomfortable questions about roles, review and how to avoid shipping poor-quality work.

At an engineering offsite, Augment ran a hackathon in which participants used agents rather than writing code themselves. The prompts were how to make an agent, a team or an individual much more effective. Perneti says the team produced substantial work and also surfaced a mix of uncertainty, frustration and fear alongside curiosity and creativity.

Instead of presenting a finished management playbook, Perneti asked teams to identify necessary tools and human behavior changes. They proposed improving agent instructions and skills and adopting specifications as a first-class way of coordinating. He sees the exercise as creating room for people to work through the transition themselves.

Berman returns to the review bottleneck: if agents produce more code than humans can inspect, how can a team understand what it ships? Perneti says Augment is adapting its review product to let agents flag the portions that need human attention. His proposed 80% automated review and possible later reduction to 1–2% human-reviewed code are aspirations, not verified safety outcomes.

On hiring, Perneti says Augment briefly paused to rethink what it should look for in an AI-native engineer and plans to reopen hiring. He stresses architecture, product judgment, user empathy and understanding an outcome rather than simply counting lines of code. A feature no one uses should prompt investigation into the user problem, not be counted as success merely because it shipped.

Pressed on whether the pause means the company can grow without adding staff, Perneti reiterates that it intends to resume hiring. He says rapid changes make an end-of-year headcount target difficult to defend, so the company is reassessing its criteria rather than assuming its old process still fits.

Perneti contrasts a discarded 2025 annual plan with a 2026 approach centered on a longer-term vision and roughly three-month execution plans. He acknowledges the uncertainty; Berman notes that engineering leaders give very different answers, from hiring aggressively to stopping recruitment.

Asked how people can understand exponential growth, Perneti tells the story of rice doubling each day. The early increases look manageable, but repeated doubling quickly leads to an enormous number. He uses it to illustrate his belief that coding-agent capability is changing faster than organizations can easily internalize.

Perneti points to the recent adoption of agents by engineering teams as a sign of rapid change and predicts that small human teams will coordinate many agents on production software. After he leaves, the hosts reflect on anxiety in the industry and on non-engineers gaining the ability to build software with agentic tools.

The second guest is Illia Polosukhin, NEAR Protocol co-founder and a co-author of Attention Is All You Need. He describes a proposed future in which personal agents protect users' interests and interact economically with other agents and businesses. NEAR's pitch combines a personal AI layer with a neutral protocol for coordination and transactions.

Polosukhin traces the project back to 2017, when he left Google to work on teaching machines to code. He says the original AI effort was early for the available compute and that the team shifted toward blockchain infrastructure before returning to AI. His shorthand is AI as a front end and blockchain as a back end.

Berman suggests blockchain could help agents transact without the same human approval flows used in conventional payments. Polosukhin argues that commerce involves more than moving money: parties must discover one another, negotiate outcomes, use escrow and resolve disputes. This is his proposed architecture, not proof that blockchain is necessary for agent transactions.

Polosukhin describes NEAR's intent-based approach, where a user specifies an outcome and a counterparty or agent can fulfill it. He gives examples ranging from currency exchange and freight to building a website or running a campaign, and says the system includes a dispute process. He cites $13 billion processed by related financial infrastructure as a company figure, without providing an independently audited measurement in the conversation.

Berman asks about crypto's reputation for scams and the reluctance of some AI communities to engage with it. Polosukhin acknowledges that permissionless systems also create opportunities for crime and says users need experience to spot scams. His view is that quality products should make the underlying technology less visible, like infrastructure behind an ordinary marketplace.

Asked whether that reputation affects adoption and hiring, Polosukhin says NEAR benefits from expertise in both AI and blockchain. He points to inference, agent hosting and confidential computing work, arguing that the two areas can serve complementary roles. These are descriptions of his organization's approach rather than a general validation of the business model.

The hosts ask what it felt like to co-author the Transformer paper. Polosukhin recalls working on question answering at Google, where recurrent models could take too long to process text for a production search setting. The team wanted a way to exploit parallel compute and relate each part of an input to the rest.

He uses the film Arrival as an analogy for processing a sequence in parallel, then describes attention as a mechanism for relating words and their context across multiple layers. The comparison is illustrative: practical Transformer systems have varying architectures and generation can still be sequential.

Polosukhin says the initial idea required experimentation to reach state-of-the-art results, and at the time many architectures were competing. The team could see meaningful progress, but the full later impact was not obvious. He left to pursue machine programming and says his timing was early relative to the compute available.

Berman turns to IronClaw and its relationship to OpenClaw. Polosukhin predicts that agent harnesses could become a new operating-system-like layer and claims browsers and apps may disappear. He frames IronClaw as a security-focused framework inspired by the emerging personal-agent workflow; the disappearance of existing interfaces is his prediction.

Polosukhin says IronClaw uses defense in depth, isolated WebAssembly tools, prompt-injection and data-exfiltration detection, and controls around credentials. He describes policy checks that restrict when secrets can be used, plus isolated subagents. These are design claims from the project's builder; they do not establish that the system is immune to attacks.

He notes that the project was only a few weeks old and still had substantial work ahead, with open-source contributors adding functionality. The hosts ask whether ordinary users will run agents and language models locally. Polosukhin expects many people to prefer hosted services, while enthusiasts will continue to experiment on their own hardware.

Polosukhin describes NEAR's confidential hosting as intended to provide an always-on service with isolated access to data and inference. He suggests a hybrid in which local models handle some simpler tasks and remote models handle others. His privacy assurances are provider claims, and practical security depends on implementation, configuration and trust boundaries.

His personal examples include morning news summaries, gathering GitHub changes into a changelog, and preparing a draft social post. A hosted agent can continue those routines while his laptop is closed. He argues that moving smoothly between local and remote resources is useful for this kind of assistant.

Polosukhin says his agent also helps synthesize objectives, bottlenecks, risks and decisions across different NEAR-related organizations. He can put the resulting summary into a shared document and use it like a chief of staff. He also describes directing several coding subagents on different features.

After the interview, the hosts discuss Polosukhin's prediction about browsers. Berman says browser control remains a frustrating obstacle in his own agent setup and that many consumer tasks depend on navigating sites whose publishers may resist automated access. Wentz expects the tooling to improve but treats that as a future possibility.

Their final guest, Modular CEO Chris Lattner, created LLVM and helped develop Swift. Asked to explain LLVM, he calls it a way for software to target different processors without every developer dealing with each chip's details. Beyond language compilers, he says the platform is used for graphics and database optimization.

Lattner describes how LLVM grew from his master's research into an open platform. Hardware vendors and other companies could collaborate on shared infrastructure even while competing commercially. He credits that community and the ability to support many architectures as major factors in its adoption.

He recalls joining Apple in 2005, when the company wanted stronger compiler and developer tools while shifting from PowerPC to Intel. Apple let the university-origin project remain open source, and Lattner worked both on Apple's needs and on a community that could support others. He later helped develop software enabling the company's silicon transition and its first 64-bit phone.

Lattner sees an echo of that earlier hardware fragmentation in today's AI market. Many companies make accelerators with different strengths, yet developers face substantial work to use alternatives. Modular's goal, he says, is to provide software, compilers and language tools that make it easier to deploy AI workloads across hardware vendors.

He argues that a thin compatibility layer or short sprint cannot resolve the whole problem: performance and developer experience require work throughout the stack. His comparison with Swift is that better tools can broaden who can build applications; Modular aims to open low-level accelerated computing to more developers while giving customers more hardware choice.

Asked why AI provokes so much anxiety, Lattner says both hype from vendors and blanket dismissal of the technology obscure the practical picture. He views AI as a tool to apply to concrete problems. He expects more software and productivity, but cautions that claims about the ultimate impact on employment need evidence and time.

Lattner says Modular had just over 170 employees and was hiring for compilers, systems, performance and inference. The company was working with advanced research organizations and had a cloud platform in beta; his planned August general-availability date was a forecast at the time of the interview, not a confirmation of a later release.

On customer needs, Lattner says organizations want reliable inference, good performance and total cost of ownership, plus the option to allocate workloads across different chips. He uses a 40% better cost example to illustrate potential savings, not a universal measured result for every deployment.

Asked what engineers Modular wants to hire, Lattner says the company considers both junior and senior candidates and values willingness to learn. He suggests that engineers directing agents must think like managers about objectives, coordination and long-term architecture, not only the code they produce.

Lattner argues that newer graduates may adapt quickly to the tools, while experienced engineers bring other strengths. He distinguishes deep systems and performance work from rapidly prototyped web applications, and says code generation is only part of software: product understanding, architecture and collaboration remain essential.

Berman notes the contrast between companies hiring juniors, pausing hiring and favoring senior staff. Lattner says businesses face different constraints and offers his opinion that some leaders may use AI narratives to justify decisions made for other reasons. His preference is a balanced team combining experience with people eager to learn.

The discussion turns to falling software stocks and the possibility of using agents to clone parts of existing products. Lattner says code is only one part of a business: data, customer relationships and support also matter. Faster implementation may accelerate experimentation, but a share-price movement alone does not show that software businesses have disappeared.

Berman asks whether software will collapse into an agent layer above a system of record. Lattner disagrees with treating that as universal: he sees agents as an additional category useful for some workflows, not a replacement for every interface or device. He urges organizations to connect the technology to real customer needs instead of adopting it solely because it is fashionable.

The hosts thank Lattner and close by reflecting on the guests' different views of engineering culture and hiring. Their conversation leaves the trajectory of jobs and agent reliability open, while showing how differently companies are already approaching the transition.