Agentic AI Crosses Into Verified Work: GPT-5.6 'Ultra' and Claude Sonnet 5 Raise the Bar for Autonomous Coding and Customer Support

Executive Summary

Frontier model providers accelerated their push into agentic AI this week, with OpenAI previewing GPT-5.6 (Sol/Terra/Luna) — including an "ultra" subagent mode — and Anthropic positioning Claude Sonnet 5 as capable of carrying multi-step coding tasks through to verified, tested completion. These are not incremental upgrades; they represent a qualitative shift in what AI systems can independently accomplish.

In enterprises, AI is increasingly deployed as a force multiplier for customer-facing and knowledge work roles. Bank of America's EricaAssist now supports over 18,000 customer service representatives in real time, while some large employers continue restructuring alongside automation — most visibly Microsoft's 4,800 job cuts. Meanwhile, a lawsuit against Meta highlights an emerging risk that cuts across every industry: when AI tools influence layoffs or performance management, workers may struggle to prove exactly how algorithmic systems were used.

New AI Tools and Model Launches

OpenAI — GPT-5.6 Series (Sol, Terra, Luna)

OpenAI released a limited preview of the GPT-5.6 model family, comprising three tiers: Sol (flagship), Terra (balanced), and Luna (fast and cost-efficient). The standout capability for workplace automation is "ultra" mode, which orchestrates multiple subagents to accelerate complex, multi-step tasks. Sol also introduces a max reasoning effort setting for deeper analytical work, and OpenAI specifically highlights strong performance on command-line tool coordination benchmarks.

The roles most directly in the crosshairs are software engineers, SRE/DevOps professionals, cybersecurity analysts, data and ML engineers, and technical PMs — particularly those in junior or intermediate positions whose work centres on routine implementation, debugging, and scripting. For senior engineering and security roles, the model functions more as an augmentation tool than a replacement.

Anthropic — Claude Sonnet 5

Anthropic's Claude Sonnet 5, now available across all plans and via Claude Code and the API, is described as the most agentic Sonnet model to date. Anthropic's own positioning is significant: the model is designed to plan, use tools such as browsers and terminals, persist across long-horizon tasks, verify its own output, and carry pull requests through to tested, merged results.

The career implications are most acute for junior software engineers, QA and test engineers, technical support specialists, and analysts performing repetitive research and synthesis. In narrow contexts — simple CRUD features, basic scripting, test generation — Sonnet 5 is capable enough to substantially reduce headcount requirements, not merely augment them.

The Broader Agentic Tooling Ecosystem

GPT-5.6 and Sonnet 5 do not exist in isolation. Across the industry, a proliferation of "managed agents" infrastructure — including Google DeepMind's Gemini API Managed Agents and agentic coding environments from other vendors — is rapidly lowering the barrier to deploying persistent AI workers. The tooling layer (background tasks, remote function calling, managed orchestration) is what converts a capable language model into a digital coworker that operates on a task queue. For professionals in roles defined by predictable, queue-based workflows, this infrastructure shift deserves close attention.

AI Adoption in Enterprises

Bank of America — AI Guidance for Customer Service Representatives

Bank of America enhanced its EricaAssist tool to deliver real-time contextual guidance — in under three seconds — to more than 18,000 customer service representatives. The system is designed to help agents resolve client needs faster and more consistently, surfacing relevant information and suggested responses during live interactions.

This deployment illustrates the dominant enterprise pattern right now: AI as a force multiplier rather than an immediate replacement. Customer service roles are not disappearing overnight, but the expectations attached to them are changing. Success in these roles will increasingly depend on exception handling, empathy, regulatory compliance, and the ability to work effectively alongside AI-generated guidance — not on procedural recall.

Microsoft — Layoffs Amid an AI Transition

Microsoft announced 4,800 job cuts, representing approximately 2.1% of its global workforce, with the largest concentration in the Xbox division (around 3,200 roles). Internal communications noted that eliminated positions are "not being replaced by AI," while simultaneously acknowledging that AI is changing how work gets done and automating routine tasks across the organisation.

The apparent tension in that messaging is itself a signal. Large technology employers are conducting structural reorganisations in parallel with AI capability investments, and the causal relationship between the two is not always made explicit. For workers and career planners, the practical implication is the same: roles defined by routine, repeatable tasks are structurally more vulnerable during periods of organisational reset.

Thomson Reuters — The "AI-Native" Workforce Shift

Thomson Reuters confirmed it is cutting a small number of engineering roles as part of an aggressive AI deployment strategy. The company's broader roadmap points toward eliminating up to 500 positions over two years while simultaneously hiring more than 250 net-new engineering roles — the large majority of which will be senior and AI-native. The legal and knowledge-work software sector is therefore showing a clear signal: the total headcount in engineering may shrink, but the seniority profile and skill expectations of those who remain will rise sharply.

AI-Resistant Skills and What Is Becoming More Valuable

High-Trust Client Advisory and Relationship Roles

As AI handles routine queries and procedural guidance — as demonstrated by BofA's EricaAssist — the differentiating value of human professionals shifts toward judgment, accountability, and regulated outcomes. In financial services, complex advice that involves fiduciary responsibility, emotional sensitivity, or ambiguous regulatory interpretation remains firmly human territory. The same pattern is emerging in healthcare, legal advisory, and any domain where accountability cannot be delegated to a system.

AI Governance, Risk, and Audit

The Meta lawsuit discussed in the following section has direct implications for which skills will be in demand. As organisations use AI in consequential decisions — hiring, performance management, layoffs — they will need professionals who can design, document, and defend those processes. AI governance and algorithmic audit is transitioning from a niche compliance function to a core organisational capability.

The Labor Market Context

Research cited by global staffing firm Adecco argues that AI is changing tasks more than it is eliminating jobs in aggregate, and calls explicitly for redesigning roles so that AI complements rather than replaces human work. Adecco's framing emphasises collaboration among companies, governments, and education systems to drive upskilling and reskilling at scale. While that systemic view is important context, individual workers cannot wait for institutional responses — the more actionable insight is that roles with clear AI-complementary components are more resilient than those defined entirely by tasks AI can now execute.

AI, Employment Law, and Worker Rights

The Meta Lawsuit: Proving AI's Role in Layoffs Is Difficult

A Reuters-reported case describes a lawsuit filed by 26 former Meta employees alleging that the company used AI tools — including productivity scoring and behavioural monitoring — in layoff selection in ways that were discriminatory. A judge in the case pointedly noted the core evidentiary challenge: the plaintiffs "were not in the rooms where it happened." Without access to the specific algorithmic criteria, model outputs, or decision logs, demonstrating a causal link between AI system outputs and discriminatory outcomes is legally arduous.

This case has broad implications beyond Meta. As AI tools become embedded in performance management, headcount planning, and workforce analytics across industries, the absence of documented decision criteria, audit trails, and appeal pathways creates risk for both workers and employers. HR professionals, employment lawyers, and people managers will increasingly need to understand how to design and evidence these governance structures — not as a legal technicality, but as a core professional competency.

Profession-by-Profession Impact

Customer Service Representatives

Bank of America's expansion of EricaAssist to 18,000+ representatives is a preview of enterprise-wide patterns. Routine troubleshooting and procedural guidance are being automated in real time; human work is shifting toward complex cases, escalation judgment, and outcomes that require empathy or regulatory accountability. Workers in these roles should expect higher throughput expectations and build skills in de-escalation, compliance scripting, and effective use of AI guidance tools — including the ability to verify, document, and escalate when AI outputs are incomplete or incorrect.

Junior Software Engineers and QA Professionals

The combination of GPT-5.6's subagent orchestration and Claude Sonnet 5's ability to carry pull requests through to verified completion represents a step-change in the automation of entry-level coding tasks. Small features, bugfixes, test generation, and repetitive refactoring are increasingly within reach of autonomous AI systems. Engineering teams will hire fewer "ticket executors" and more professionals capable of systems thinking, production debugging, security review, and designing evaluation frameworks for AI-generated code changes.

DevOps, SRE, and IT Automation Professionals

OpenAI's emphasis on command-line workflow performance and long-horizon security tasks — combined with subagent mode for autonomous remediation — signals increasing automation of runbooks and routine incident response. The human SRE's value shifts toward reliability engineering at the system level: designing guardrails, building safe automation pipelines with appropriate approval and rollback mechanisms, and leading postmortem-driven architectural improvements. Specialising in observability and policy-as-code will become more defensible than executing standardised playbooks.

HR Professionals, People Operations, and Employment Lawyers

The Meta lawsuit crystallises a trend that has been building for several years. As AI tools become embedded in performance management and workforce decisions, HR and legal professionals who can bridge the gap between technical systems, HR process design, and legal risk will be in high demand. The practical skills gap to close: algorithmic audit basics, documentation standards for AI-influenced decisions, vendor risk assessment, and the design of internal "AI decision logs" that can withstand legal scrutiny.

Key Takeaways for Career Strategy

• Agentic AI is now verified, not theoretical. Both GPT-5.6 and Claude Sonnet 5 can complete multi-step technical tasks end-to-end. Roles defined by task execution within a queue are structurally more exposed than roles defined by judgment, design, or accountability.
• Enterprise adoption is accelerating in customer-facing roles. BofA's 18,000-rep deployment is not an experiment — it is production infrastructure. Customer service professionals should treat AI tool fluency as a baseline competency, not an optional skill.
• The demand profile for engineering is bifurcating. Thomson Reuters' hiring plan — fewer total engineers, but a higher proportion of senior and AI-native roles — is a visible expression of a trend playing out across knowledge-work industries.
• AI governance is becoming a career-defining skill. The Meta lawsuit is an early indicator of the legal and operational complexity that follows when AI systems are used in consequential employment decisions without adequate documentation. Professionals who can design, audit, and explain those processes will have durable value.
• Document your work and decision criteria now. Regardless of role, building personal audit trails — of performance reviews, decision rationales, and the criteria applied to you — is a practical risk-management step in an environment where AI involvement in employment decisions may be difficult to surface or challenge after the fact.