AI transforms how work gets done

AI is driving a fundamental shift in software applications from consumption of standalone models as chatbots to multi-agent systems that can plan, act, and adapt in real time. However, in 2025, the agent landscape is the wild west. The process of building and deploying agents is hit or miss, undisciplined, and exposes businesses to new kinds of security vulnerabilities if agents are not built with the greatest of care. It is possible to create stellar agents, but it is neither easy, nor guaranteed, and it is expensive.

Hugging Face CEO Clem Delangue once said "when it works, it's software." In 2026, generative AI will "grow up" and the process of building and deploying AI will become more disciplined, more orderly, safer, and (in a good way) a little more boring. Middleware and frameworks will emerge that make agent building less art and more science, resulting in more predictable outcomes.

At IBM, we are pursuing an ambitious research agenda under the banner of "generative computing"--the idea that generative AI can be woven into software systems, and that deterministic traditional software and generative AI can not only coexist but can fundamentally complement each other's strengths and weaknesses. In generative computing, agents are just one example of generative programs, and new programming models will emerge to make usage of AI more predictable and secure.

The capabilities of models will continue to improve, following a Moore's Law-like progression where every 9-12 months we see a 10-fold reduction in the size of a model required to achieve a certain level of capability. This will make AI increasingly ubiquitous. We will see a broader range of software quietly using AI, often in less visible but equally powerful ways, as AI becomes commonplace enough to suffuse software of all kinds.

We expect all major IBM products to adopt agentic systems capable of reasoning and reaching into enterprise backend systems. For instance, in the context of data systems, agents will begin to autonomously tackle data systems design and operations tasks currently performed by engineers and analysts. For example, autonomously designing and validating novel data products based on demand, generating dataflow pipelines to discover, index, catalog, clean, and validate data for consumption by generative AI, and remediating dataflow issues as workload or resources evolve. Data systems will evolve to better address agentic data access patterns, which will be more iterative than current OLAP/OLTP queries, as the agents proactively assist users in refining their intent and the data to achieve their analytic goals. To address the cost overhead associated with data processing at enterprise scale, agents used at design time will produce efficient flows that minimize model and agent use at flow execution time.

Entire ecosystems of AgentOps will emerge, focusing on monitoring, debugging, and orchestrating agents at scale. Enterprises have embraced a hybrid platform and infrastructure strategy as fundamental to their IT strategy. Hybrid AI systems will integrate into hybrid applications, platforms, and infrastructure as agents will increasingly integrate with existing ecosystems via tools. To guarantee operational control, AgentOps will be integrated in existing observability and operations strategies and solutions.

Security concerns, including agent hijacking, prompt injection risks, credential theft, excessive permissions, tool manipulation, and insufficient monitoring and access controls will be mitigated via the use of ephemeral agent identities, just-in-time tokens, and delegation frameworks. Sandboxing, red/blue testing, and real-time policy enforcement will provide autonomous defense.

As AI becomes more woven into the fabric of software, the lines between software and AI will blur. Developers already use AI to create software today, but increasingly, AI will write software--including software that uses AI. AI will generate programs to solve problems better suited to traditional software, but it will also generate software to orchestrate its own operation, allowing it to overcome some of its own inherent weaknesses, accelerating progress.

Application development will take new forms. Development might start with the specification of an agent, but that agent will produce code as it operates, leaving behind a trail of deterministic software as it charts a path to meeting the applications needs. As an application matures, the agent will only resurface when corner cases are goals, patching and hardening code to fill gaps. Human developers and AI will collaborate on the creation of applications, but the balance will shift over time until most human developers play an increasingly high-level supervisory role, along the lines of a product manager. Average workers will increasingly produce ephemeral automations as a routine part of their work, though it won't seem any less natural than using a spreadsheet.

Legacy software will be tended by AI systems that first encircle these legacy software systems and eventually become a part of them (though never fully managing to eradicate them). Agent-to-agent communications will move beyond being novelties and will be become the de facto way that IT systems interact with one another.

We have already seen AI perform at the level of a silver medalist in the International Mathematical Olympiad; address benchmark problems in math, science, and programming; help find conjectures and rule out counterexamples; check proofs; and power agents to improve pieces of algorithms by changing code. We will see this trend gain steam as these tools are further developed to be used to guide human intuition when tackling science and engineering problems.

Security will become a sophisticated arms race, with new attack surfaces, new insider threats, and the equivalent of social engineering attacks, but aimed at AI systems, will become commonplace. The agent landscape will become a battlefield of increasingly active countermeasures, and some critical systems might begin to exclude the use of AI or even be partially air-gapped to protect them, forming a paradoxical retrograde trend.

As multi-agent systems become prevalent in the enterprise, we will provide advances in agent-to-agent authentication and encrypted inter-agent protocols.

Multi-agent observability will track issues of multi-agent collusion and coordination attacks, providing anomaly detection and consensus validation. Agents will continue to operate in existing enterprise ecosystems, hence the need for true hybrid operations and observability will remain. Agents will determine their own observability needs, based on feedback from the environment that they operate in, and auto-instrument themselves for operational visibility and action.

Agents with pervasive observability and memory will lead to a much more dynamic and optimized environment for data which blurs the line between data and APIs. Agents will use this memory to learn the user's needs and will prepare the needed data in a just-in-time approach, minimizing effort expended. Further, the end user will be unaware of whether data is coming from a data repository or an API as agents autonomously determine the best way to provide users with needed information.

As AI subsumes IT, the nature of work and society will change, as it has in previous waves of technological change, from the invention of agriculture, to the mechanization of physical labor, to the rise of the internet.

The mix of jobs undertaken in the economy will have begun a transformation by this point, with associated disruption of social orders and an evolution of public institutions, as in previous waves of technological change. AI will be an inextricable part of almost every occupation, just as the internet and telecommunications today are a part of practically every job, but human workers will remain as important as ever, despite futuristic predictions of singularity and the full replacement of human labor. Ubiquitous physical robots, perennially "just around the corner," will gradually become more common and cost effective beyond the niches they occupy today, though their development cycles will remain sluggish relative to the progress of AI.

Significant fractions of business operations will be automated (or at least, automatable), though legacy systems will remain difficult to fully shake free of, and availability of energy and natural resources will serve as a brake on an otherwise brisk technological acceleration (barring a major energy breakthrough, which seems unlikely).

AI tools will increasingly become a part of scientific discovery, accelerating progress in many fields, but that progress will disappoint those holding more exuberant techno-utopian visions. Science and invention will remain slow, methodical endeavors, albeit empowered with new tools, as in past waves of innovation.

As physical AI or robotics emerge, become more prevalent, and interact with existing enterprise systems depending on their specialty, there will be platforms to observe and govern them, integrating across multiple physical AI vendors.

As AI becomes increasingly capable of tackling tasks that previously resisted automation due to inherent complexities, it will open the door to fully autonomous data systems. That is, multi-agent systems will build entire, self-sufficient data stacks from scratch, ranging from data infrastructure design, to data discovery, integration, governance, and even proactively pinpointing relevant insights. Towards this goal, human interventions, both at design time and ongoing operations, will focus on clarification of goals, approval for state-changing actions, and receiving results.

A distributed platform enables hybrid computing

The AI innovation shift from models to AI systems that incorporate models with traditional components, and the applications built on top of these systems, represent a novel paradigm. Standards are emerging around these new concepts and interactions. We can view these trends as the beginning of a new application and protocol layer (layer 8) that leads to the evolution of middleware and platform. It creates a need to build an "operating system" for AI (AIOS) to manage and provide common services to this class of applications.

AI innovation also intensifies the need for operating such non-deterministic systems at scale while integrating them into existing ecosystems. A wave of new security threats is also emerging stemming from cloud-specific agent indeterminacy in distributed environments. These create the need for novel observability and security mechanisms.

The AIOS will support development, life-cycle management, and operations of AI workloads across the entire stack: models, tools, persistence, and applications. AIOS for enterprises will be implemented by extending and adapting an existing cloud platform rather than by building a new stack, as this allows the reuse of capabilities and best practices that provide reliability, security, and compliance.

We will deliver a hardened, enterprise-grade inference stack for OpenShift AI with a production-ready control plane, SLO-driven autoscaler, pluggable scheduler, fast model actuation, and multi-tiered kv-cache offloading for heterogeneous hardware. We will develop a simplified and consistent experience for connecting models to data and will consistently and flexibly scale AI across hybrid cloud. Storage with AI capabilities (content-aware storage) will enable storage to process data effectively for agentic AI applications. We will have an MCP platform with registry and life-cycle management; an MCP gateway for agent-tool discovery, connectivity, and access control; and MCP servers for agents to manage hybrid cloud platform and infrastructure.

Focused AI agents will assist with system operation by generating recommendations and remediations that build on top no/low-code techniques, supporting humans in the loop for change request approvals. AI agents will be able to define (code) development and operations guard-rails, observe their effectiveness in live systems autonomously, and fine-tune as needed. The ability to continuously and intelligently assess applications and operations against benchmark expectations will yield richer visibility into operations risk and deliver higher trust and confidence in AI agents in operations.

We will deliver trusted agent identity support for agent-tool authorization flows based on existing and emerging standards. We will implement best-of-class security by integrating IBM enterprise security tools with AIOS to achieve identity management for agents; data protection; security posture management; and governance, risk, and compliance.

Data-centric workloads will drive the need for batch inference using smaller-scale models to support various ingest and analytic flows in which millions of records or documents need to be processed. These scales also have significant cost implications, further driving the need for efficient bulk inference with bursty demand. Finally, the need for low-latency inference associated with data access will continue to drive the push of inference into data repositories.

Our integrated AI agent platform will have agent life-cycle and layer-8 interconnect for security, safety, metering, quality enforcement, and observability, including seamless interactions of humans into agent life-cycle and operations. The llm-d "inference brain" will provide core scheduling, kv cache management, and autoscaling logic behind platform-agnostic APIs with key enterprise features for security and multi-tenancy.

It will integrate with AI hardware like AIU Spyre. KV-cache-optimized storage will improve inference performance and cost.

Return on AI investment will depend on the ability to specialize AI capabilities to specific domains and enterprise workflows. For this, models and inferencing platforms will need to be customized to enterprise needs, AI platforms integrated with existing processes and systems, and AI agents will need to learn from feedback and adapt to context. Use of multiple models will become common as a way to optimize quality, performance, and cost for specific use cases. Agent memory will emerge as a new layer of the data stack. To accomplish complex tasks, agents will interact with one another without human coordination, leading to adoption of multi-agent systems (MAS). We expect further evolution of standards, focusing on reliability, security, observability, and trust. With more adoption of standards, ecosystems for tools, agents, evals, and others will expand.

Production deployments will need to address concerns of infrastructure-level indeterminacy and security while adhering to compliance regimes. For instance, the scaling attack surface will expand across hybrid infrastructure, and manual operations will be overwhelmed by the distributed cloud complexity.

DevOps practices and tools will change to incorporate the increasing use of coding agents without compromising software reliability, security, and performance. Autonomous, self-healing systems will emerge, and root cause detection will be more accurate and independent. Coding agents will provide autonomous remediation.

Cloud platforms will be more frequently built by local vendors in small-medium size data centers driven by sovereignty concerns. AIOS will support multi-tenant deployments backed by multi-tenant infrastructure and platform packaged software. It will be operated by third-parties.

In 2028, we will build a fully-formed distributed operating system for AI inference in the era of agentic workloads. We will showcase a unified control plane that federates and orchestrates entire graphs of multi-modal agents across a hybrid fabric, becoming the foundational substrate for next-generation AI applications.

We will develop multi-party protocols for hybrid environments and automated policy management and compliance across clouds (policy as code). We will provide real-time monitoring, evidence collection, threat correlation, incident orchestration, distributed audit, and AI/ML analytics; and implement infrastructure-level consensus mechanisms.

We will also lay a foundation for interoperability at scale for MAS, with a visibility and control layer across all interactions in MAS and tools to diagnose unexpected outcomes of AI workflows, assign accountability, and govern and audit MAS behavior. We will develop interfaces and workflows to integrate MAS with humans and systems into seamless multi-modal interactions. There will be persistent memory to capture and share an agent state and context. We will reinvent processes for systems and application management with infrastructure automation leveraging AI in a trustworthy and reusable manner. There will be tools to support agents and secure their life-cycle.

Management and observability agents will handle reliability engineering, operational risk, security, compliance, and business impact assessments, and will provide recommendations based on business and technical KPIs.

We will deliver Fusion storage with support for hybrid cloud multi-agents (kv cache and vectors sharing), support for agentic AI across structured and unstructured data, and integrated storage management into MAS for autonomous storage operation.

We expect the continued expansion of AI agents and the emergence of practically useful quantum computing to drive the technology landscape after 2028.

The continued transformation of industries with AI will lead to a decentralized world-wide system of agents interacting in a variety of patterns. As was the case with the internet, this will require further evolution of protocols, standards, and control planes to support discovery and connectivity. Enterprises will require new mechanisms to control the information flow, prevent security exposures (including novel ones), and ensure alignment with organization objectives regardless of uncertainty.

This agents web will drive a multi-fold increase of compute demand per knowledge worker across all aspects of AI: models, applications, state management, and operational management.

This demand will likely lead to diversification of hardware specialized on specific tasks. If this holds, infrastructure will become even more hybrid, with emphasis on open architectures with open standard abstractions across the entire stack.

The evolution of LLMs will significantly slow down and training costs will dramatically decrease. This will enable the creation of truly competitive open weights models. This may lead to the standardization of prompt formats, customization methods, and drive faster innovation in the application layer.

As quantum computing matures to practical applicability, enterprise use cases will also emerge. Innovative organizations will want to integrate quantum computing capabilities into their applications and workflows, leveraging both AI and quantum computing for enterprise problems. We expect to see and influence the emergence of new applications that leverage high precision computing, AI, and quantum computing. Particularly in an enterprise context, leveraging these technologies will require improved alignment between the domains of enterprise application and high-performance computing.

IBM infrastructure will support a quantum computing and agentic AI platform with a single storage/data management.

Meanwhile, large-scale fault-tolerant quantum computers threaten to break hybrid cloud cryptography. This is addressed with post-quantum cryptography for hybrid infrastructure. Other security risks include potential cross-cloud data exposure from quantum computing advances, and infrastructure-level indeterminacy in a world of parallel execution and quantum workflows. This will be addressed with confidential computing, decentralized NHI permissions, and leveraging uncertainty for infrastructure verification.

The future of computing is quantum-centric

We will have the first examples of quantum advantage using a quantum computer with an HPC. This integration of heterogeneous systems will lead to the development of a standard to allow seamless integration with classical systems across different quantum hardware vendors. Users will be able to run workloads that involve quantum and classical resources by writing quantum and classical code and deploying it in an integrated system. In 2026 and beyond, we will introduce new profiling tools to help users monitor, verify, and debug workloads across quantum and classical resources.

IBM Quantum Nighthawk is our platform for exploring and scaling quantum advantage ahead of large-scale fault-tolerant quantum computing. It uses a square lattice, connecting each qubit to up to four neighbors. Paired with techniques that reduce errors, Nighthawk is expected to run circuits with 7,500 gates in 2026 in with up to three 120-qubit modules (360 qubits), 10,000 gates in 2027, and 15,000 gates in 2028. As we improve and scale Nighthawk, we expect major progress toward quantum advantage.

IBM Quantum Loon debuted in 2025 with a new chip architecture that leverages c-couplers to link qubits across the chip, beyond nearest neighbors. It enables up to six degrees of connectivity between qubits. This improved connectivity is needed to implement IBM's scalable error-correcting code. This advancement strengthens our confidence in achieving large-scale, fault-tolerant quantum computing by 2029. In 2026, we will prototype our error correction decoder, which will enable real-time error correction--a key capability for scalable, fault-tolerant quantum computing.

We will continue to deliver improvements to dynamic circuits, error mitigation, and speed in the coming years to extend the capabilities of our available quantum systems.

We will introduce utility mapping tools that support the exploration and design of new algorithms by mapping problems to circuits that scale prior to fault tolerance. We will also work with partners to create a use case benchmarking tool, enabling others to explore which of their applications are ripe for near-term quantum value.

Patterns will begin to appear in workflows within the quantum advantage regime, providing the opportunity to start applying AI-driven automation to combine quantum and classical resources and set up and manage the components and configurations of these workflows. AI will also start driving improvements in quantum system setup and enhance the developer experience by enabling more efficient code development.

To prepare for the future, Kookaburra will demonstrate a single module out of Starling consisting of a logical processing unit and quantum memory.

Meanwhile, as the technology progresses toward large-scale, fault-tolerant quantum computing, the need for quantum-safe cryptographic protocols to protect data in flight and digital signatures intensifies. Cryptographically-relevant quantum computers may not arrive for some years but completing an inventory of cryptography in an organization, performing a risk assessment, migrating vulnerable cryptography, and implementing a crypto-agility framework will take several years. This, combined with harvest now, decrypt later threats, means enterprises needing strong data protection should already be evaluating and implementing an action plan.

In 2027, we will diversify quantum advantage and entangle fault-tolerant modules. The performance of our Nighthawk processor will improve to allow circuits with up to 10,000 gates on up to 1080 qubits in 2027, scaling to up to 15,000 gates in 2028.

We will introduce computation libraries. These provide mathematical subroutines for applications that integrate with popular existing computational libraries to support collaboration, integration, and help users more efficiently build and orchestrate workflows across classical and quantum compute resources.

We will introduce workflow accelerators that deliver optimized quantum-classical execution pipelines for efficiently running similar tasks. This will save developers valuable compute time.

In 2028, we will prototype a complete instruction set architecture including magic state distillation for fault-tolerant quantum computing in our Starling proof-of-concept, to be released in 2029.

To prepare for the future, in 2027 the Cockatoo processor will demonstrate entanglement of two modules using a universal adapter. In 2028, we will demonstrate multiple modules and magic state distillation.

Fueled by the momentum of quantum advantage, we expect to see agentic tools emerge that efficiently map problems to quantum workflows to help domain experts and industry practitioners engage more easily and effectively with quantum computing.

The first fault-tolerant quantum computer, Starling, will be available to clients in 2029. It will be a modular, error-corrected quantum-centric supercomputer with 200 qubits capable of running 100 million gates. It will enable the development of more sophisticated circuit libraries. We will continue to scale electronics, infrastructure, and software to reduce footprint, cost, and energy usage.

We expect to provide libraries that offer an expanded, generalized set of circuits ideal for execution on fault-tolerant processors.

By 2033+, we will scale fault-tolerant quantum computers to deliver a system, called Blue Jay, capable of running circuits of 1 billion gates on up to 2000 qubits with a power consumption of 2 megawatts. This will require a new control electronics and cryogenics infrastructure.

Extensions of the computation and circuit libraries will scale and diversify quantum computing + HPC workflows across industries. For the future, we will scale beyond Blue Jay with the development of distributed quantum computing, bringing together the fields of quantum communication and quantum computation. These large-scale fault-tolerant quantum computers will unlock a new era of algorithmic complexity and application discovery. Developers will not need to change how they write quantum programs in this era. They will simply notice that they can run longer workloads.

The vast amount of data that these systems will process will introduce significant complexities. Mechanisms will be required to efficiently compress and stream data across different computational components. In this context, AI can play a critical role by optimizing and compressing data transfers between quantum and classical resources, ensuring seamless and efficient movement of information.

In scenarios where multi-agent systems manage entire workflows and automatically adapt considering both the problem and available resources, quantum systems will leverage this adaptability to align with the application's top layer. In heterogeneous computing environments, this will become a key differentiator, evolving agentic components into self-sufficient operational entities that support discovery, integration, and governance.