Uncategorized

Large language models become static after pretraining. Their knowledge does not update as the world changes. Retraining a full LLM is too expensive at modern scales. Fine-tuning risks degrading previously learned knowledge. Retrieval-augmented generation (RAG) struggles when answers require reasoning across many documents. A team of researchers from the National University of Singapore, MIT CSAIL, A*STAR, and the Singapore-MIT Alliance for Research and Technology (SMART) proposes a new approach called MEMO (Memory as a Model). What Problem Does MEMO Solve? Existing methods for integrating new knowledge into LLMs fall into three categories. Non-parametric methods like RAG retrieve documents at inference time. They are sensitive to retrieval noise and struggle with cross-document reasoning. Parametric methods such as continual pretraining or supervised fine-tuning internalize knowledge into model weights. They are computationally expensive and cause catastrophic forgetting, where new training degrades previously acquired knowledge. Latent memory methods compress knowledge into soft tokens. These representations are tightly bound to the model that produced them — a limitation the research team calls representation coupling which limits transferability across LLMs. MEMORY as a Separate Model MEMO separates memory from reasoning. The MEMORY model is a small, dedicated language model trained to internalize knowledge from a target corpus. The EXECUTIVE model is the main LLM — frozen and queried only through its standard input-output interface. In experiments, the MEMORY model is Qwen2.5-14B-Instruct. The EXECUTIVE model is either Qwen2.5-32B-Instruct or Gemini-3-Flash, a proprietary closed-source model. Because MEMO treats the EXECUTIVE model as a black box, it does not require weight access or output logits. https://arxiv.org/pdf/2605.15156 How the MEMORY Model is Trained Training begins with a five-step data synthesis pipeline guided by a GENERATOR model — Qwen2.5-32B-Instruct in experiments. The pipeline converts a raw document corpus into a reflection QA dataset: question-answer pairs that represent corpus knowledge under diverse query variations. The five steps are: Fact extraction — direct extraction of explicitly stated facts, and indirect extraction of inferred information, run in parallel per document chunk. Consolidation — QA pairs sharing a common context (entity, time period, relationship) are merged into multi-fact pairs. Verification and rewriting — each QA pair is checked for self-containment. Pairs with unresolved pronouns or implicit references are rewritten using the source chunk or discarded. Entity surfacing — QA pairs are generated where questions encode entity attributes and relationships, and answers reveal entity identities. This targets the reversal curse, where models trained on “A is B” fail to infer “B is A.” Cross-document synthesis — the GENERATOR model constructs QA pairs spanning multiple documents. It identifies two types of cross-document connections: converging clues (multiple documents about the same entity) and parallel properties (different entities sharing a common attribute or role). Step-5 is the most critical component. A leave-one-out ablation shows that removing it drops accuracy from 24.00% to 6.37% on NarrativeQA. It is also the dominant source of training pairs in the final dataset. The MEMORY model is then trained via supervised fine-tuning (SFT). The loss is computed over answer tokens only. Source documents are never provided at inference. The model must answer from internalized parametric knowledge. Inference: The Structured Multi-Turn Protocol At inference, the EXECUTIVE model queries the MEMORY model through a structured multi-turn protocol with three sequential stages. Stage 1: Grounding. The EXECUTIVE model decomposes the query into atomic sub-questions. Each targets a single identifying constraint. The MEMORY model answers each independently. Stage 2: Entity identification. Using the grounding responses, the EXECUTIVE model issues targeted follow-up sub-queries. It iteratively narrows down candidate entities until one is confirmed or the stage budget runs out. Stage 3: Answer seeking and synthesis. Conditioned on the identified entity, the EXECUTIVE model queries the MEMORY model for supporting facts. It then synthesizes all retrieved responses into a final answer. The MEMORY model’s responses are compact natural-language snippets. Their length is independent of corpus size, so retrieval cost does not scale with the number of documents. This contrasts with RAG, where inference cost grows with the corpus. Experimental Results MEMO is evaluated on three benchmarks: BrowseComp-Plus (multi-hop deep-research), NarrativeQA (discourse understanding over books and movie scripts), and MuSiQue (2–4 hop reasoning over Wikipedia paragraphs). Baselines include BM25, NV-Embed-V2, HippoRAG2, and Cartridges. Cartridges requires white-box access to the EXECUTIVE model and scored 0.00% on BrowseComp-Plus and 3.75% on NarrativeQA. On NarrativeQA with Gemini-3-Flash, MEMO achieves 53.58%. HippoRAG2 reaches 23.21% on the same setup. On MuSiQue, MEMO achieves 60.20% against HippoRAG2’s 57.00%. On BrowseComp-Plus, MEMO achieves 66.67% against HippoRAG2’s 66.33%. With Qwen2.5-32B-Instruct as EXECUTIVE model, MEMO achieves 54.22% on BrowseComp-Plus and 48.30% on MuSiQue. Switching to Gemini-3-Flash yields gains of 12.45%, 26.73%, and 11.90% on the three benchmarks. The MEMORY model is not retrained when the EXECUTIVE model changes. Robustness to retrieval noise: The research team evaluates performance when distractor documents are added to the corpus. NV-Embed-V2 and HippoRAG2 drop by up to 6.22% on BrowseComp-Plus when one negative document is added per evidence document. MEMO’s accuracy on the same benchmark changes by +0.55% — within one standard deviation. MEMORY model architecture robustness: The research team also tests three MEMORY model families at similar parameter scale: Qwen2.5-1.5B-Instruct, Gemma3-1B-IT, and LFM2.5-1.2B-Instruct (a hybrid state-space and transformer architecture). Performance is largely consistent across all three, indicating the framework is not sensitive to the specific pretraining lineage of MEMORY model. Continual Knowledge Integration via Model Merging MEMO supports incremental knowledge updates through model merging. When a new corpus arrives, a separate MEMORY model is trained on it independently. Its task vector — the parameter difference from the base model — is then merged with the existing MEMORY model in parameter space. The research team test this on NarrativeQA using TIES merging (ρ=0.3). For K=2 corpora, merging accumulates 48 GPU-hours versus 72 GPU-hours for full retraining — a 33% reduction. At K=10, merging scales as Θ(K) while full retraining scales as Θ(K²), yielding a 5.5× saving (240 vs. 1,320 GPU-hours). The merged MEMORY model trails full retraining by 11.04% under Qwen2.5-32B-Instruct (15.81% vs. 26.85%). It trails by 19.11% under Gemini-3-Flash (34.47% vs. 53.58%). Despite

Speculative decoding is a technique for speeding up large language model inference. A small, fast draft model proposes several tokens. The large target model verifies them in parallel. If accepted, inference is faster. If rejected, the system falls back gracefully. EAGLE Team, vLLM Team, and TorchSpec Team has launched the EAGLE series including EAGLE 1, EAGLE 2, and EAGLE 3 has become one of the most widely adopted and practically deployed families of speculative decoding algorithms across both research and production systems. Today, that family gets a targeted reliability upgrade with introduction of EAGLE 3.1. What was Going Wrong While speculative decoding performs well in controlled settings, performance often degrades under different chat templates, long-context inputs, or out-of-distribution system prompts. The EAGLE team traced this fragility to a phenomenon called attention drift as speculation depth increases, the drafter gradually shifts attention away from sink tokens and toward its own generated tokens. In simpler terms: the drafter is a small model that predicts future tokens. As speculation gets deeper, it starts attending to its own prior outputs instead of the original context. This degrades acceptance length and output stability. Two underlying issues were identified. First, the fused input representation becomes increasingly imbalanced as higher-layer hidden states dominate the drafter input. Second, hidden-state magnitude grows across speculation steps due to the unnormalized residual path. Together, these effects make the drafter progressively less stable at deeper speculation depths. Two Architectural Fixes in EAGLE 3.1 To address attention drift, EAGLE 3.1 comes with two key architectural improvements: FC normalization after each target hidden state and before the FC layer, and feeding post-norm hidden states into the next decoding step. FC normalization stabilizes the hidden states that the drafter receives from the target model. Without it, hidden-state magnitude grows across steps, making the drafter increasingly unreliable. Applying normalization at each step keeps the inputs bounded. The post-norm design makes the method behave more like recursively invoking the drafter across decoding steps, rather than simply appending additional layers to the target model. https://vllm.ai/blog/2026-05-26-eagle-3-1 What These Fixes Deliver Compared with EAGLE 3, EAGLE 3.1 demonstrates: better training-time to inference-time extrapolation, stronger long-context robustness, higher resilience to chat template and system prompt variation, and more stable acceptance length across diverse serving environments. In long-context workloads, EAGLE 3.1 achieves up to 2× longer acceptance length compared with EAGLE 3. Training Infrastructure: TorchSpec TorchSpec now provides efficient training support for EAGLE 3.1 and future speculative decoding algorithms. By lowering training overhead and simplifying experimentation workflows, TorchSpec helps accelerate iteration and exploration for next-generation speculative decoding research and deployment. Based on TorchSpec and vLLM, the research team also trained and open-sourced an EAGLE 3.1 draft model for Kimi K2.6, available on HuggingFace. The model serves as an example of deploying EAGLE 3.1 with TorchSpec training and vLLM serving support on a real-world serving model vLLM Integration: Config-Driven and Backward-Compatible EAGLE 3.1 lands in vLLM as a config-driven extension of the existing EAGLE 3 implementation. The integration includes FC normalization support, post-norm hidden-state feedback, and removal of hardcoded assumptions around target hidden states. Backward compatibility with existing EAGLE 3 checkpoints is fully preserved. EAGLE 3.1 draft models can be plugged directly through the same speculative-decoding code path. Copy CodeCopiedUse a different Browser vllm serve nvidia/Kimi-K2.6-NVFP4 –trust-remote-code –tensor-parallel-size 4 –tool-call-parser kimi_k2 –enable-auto-tool-choice –reasoning-parser kimi_k2 –attention-backend tokenspeed_mla –speculative-config ‘{“model”:”lightseekorg/kimi-k2.6-eagle3.1-mla”,”method”:”eagle3″,”num_speculative_tokens”:3}’ –language-model-only Benchmark Results on Kimi K2.6 The research team benchmarked the Kimi K2.6 EAGLE 3.1 draft model on Kimi-K2.6-NVFP4 with vLLM (TP=4, GB200, non-disagg) on the SPEED-Bench coding dataset. EAGLE 3.1 delivers 2.03× higher per-user output throughput at concurrency 1. The speedup stays meaningful as concurrency scales: 1.71× at C=4 and 1.66× at C=16. Marktechpost’s Visual Explainer 01 / 07 vLLM · May 26, 2026 Meet EAGLE 3.1 The EAGLE team, vLLM team, and TorchSpec team jointly released EAGLE 3.1 — a targeted fix for speculative decoding instability in production LLM serving. #speculative-decoding #vLLM #LLM inference #performance 02 / 07 Background What is Speculative Decoding? A technique for speeding up LLM inference using two models working together. A small, fast draft model proposes several tokens ahead The large target model verifies all proposed tokens in one pass Accepted tokens are kept — rejected tokens fall back gracefully Result: higher output throughput with no change in output quality 03 / 07 The Problem Attention Drift in EAGLE 3 EAGLE 3 performance degraded in real-world deployments under three conditions: Different chat templates Long-context inputs Out-of-distribution system prompts Root cause: attention drift — as speculation depth increases, the drafter shifts attention away from sink tokens toward its own generated tokens. 04 / 07 Root Cause Two Underlying Issues The fused input representation becomes increasingly imbalanced — higher-layer hidden states dominate the drafter input Hidden-state magnitude grows across speculation steps due to the unnormalized residual path Together, these make the drafter progressively less stable at deeper speculation depths 05 / 07 Architecture Two Architectural Fixes Fix 1 FC normalization applied after each target hidden state and before the FC layer. Keeps hidden-state magnitude bounded across decoding steps. Fix 2 Post-norm hidden-state feedback — normalized hidden states fed into the next decoding step, making the drafter behave like recursive invocation rather than appended layers. 06 / 07 Benchmarks · SPEED-Bench Coding · GB200 TP=4 Per-User Throughput vs. No-Spec Baseline 2.03×Concurrency 1 1.71×Concurrency 4 1.66×Concurrency 16 In long-context workloads, EAGLE 3.1 achieves up to 2× longer acceptance length compared with EAGLE 3. Tested on Kimi-K2.6-NVFP4 with vLLM. 07 / 07 Deployment · vLLM v0.22.0 How to Deploy EAGLE 3.1 Backward-compatible with EAGLE 3 checkpoints. Already merged in vLLM main. Stable release: v0.22.0. vllm serve nvidia/Kimi-K2.6-NVFP4 –trust-remote-code –tensor-parallel-size 4 –tool-call-parser kimi_k2 –enable-auto-tool-choice –reasoning-parser kimi_k2 –attention-backend tokenspeed_mla –speculative-config ‘{“model”:”lightseekorg/kimi-k2.6-eagle3.1-mla”, “method”:”eagle3″, “num_speculative_tokens”:3}’ –language-model-only ← Prev 1 / 7 Next → Marktechpost AI & ML Research, Simplified. Key Takeaways EAGLE 3.1 fixes attention drift — a newly identified instability where the drafter loses focus on sink tokens at deeper speculation depths. Two architectural changes — FC normalization and post-norm hidden-state feedback — stabilize the

When large language models, or LLMs for short, produce outputs, several criteria are at stake, including not only overall response relevance but also coherence and creativity.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Stay on top of what’s going on in AI this summer Here at MIT Technology Review, we understand exactly how relentless the pace of news from the world of artificial intelligence feels. New models and capabilities crop up as fast as we can cover them, and the ripple effects they send through tech and wider society are never far behind. Our unique strength lies in cutting through the day-to-day noise to help you understand what’s really happening, and what lies around the corner. That’s why we created our list of 10 Things That Matter in AI Right Now, unveiled at our flagship AI event EmTech AI a few weeks back (check the list out if you haven’t already!) And it’s why we publish so many stories dedicated to explaining how AI works, and what’s coming next. We also regularly run live subscriber-only Roundtables events—you can still catch up on last week’s session, where we explored how AI might enter the physical realm via world models. Right now, there’s a 25% discount on subscriptions. Sign up now to deepen your understanding of AI this summer. You can also join the conversation by subscribing to The Algorithm, our free weekly newsletter all about the latest in AI. MIT Technology Review Narrated: what’s next for IVF IVF has brought millions of babies into the world over the last four decades. But the process can still be slow, painful, and expensive—and far from guaranteed to work. Now, a wave of new technologies aims to change that.  Researchers are using AI to identify promising sperm and embryos, developing robotic systems that could automate parts of the IVF process, and even exploring controversial genetic editing techniques designed to prevent inherited disease. The technologies could make IVF more effective and accessible. But they’re also raising difficult ethical questions about how far reproductive medicine should go. —Jessica Hamzelou This is our latest story to be turned into an MIT Technology Review Narrated podcast, which we publish each week on Spotify and Apple Podcasts. Just navigate to MIT Technology Review Narrated on either platform, and follow us to get all our new content as it’s released. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 NASA unveiled plans for three uncrewed missions to the Moon this yearThey’re part of preparations for a crewed landing in 2028. (The Verge)+ And steps to build the first lunar base at the Moon’s south pole. (NBC News)+ Jeff Bezos’s Blue Origin will lead the first uncrewed mission. (WP $)+ NASA is building the first nuclear reactor-powered spacecraft. (MIT Technology Review) 2 Samsung’s largest unions have approved a landmark bonus schemeThe deal averts a massive strike at the world’s largest memory-chip maker. (WSJ $)+ Chip workers will get an average bonus of about $340,000. (Bloomberg $)+ The dispute centered on who profits from the AI boom. (BI)+ Resistance to AI is growing. (MIT Technology Review) 3 Elon Musk accused the Pentagon of misusing Starlink for dronesHe says military use of the system violates SpaceX rules. (Ars Technica)+ The DoD is disputing a Starlink price hike during the Iran war. (Reuters $)+ Stratospheric internet could take off this year. (MIT Technology Review) 4 China has overhauled the world’s biggest surveillance network with AIBeijing is pushing law enforcement towards predictive policing. (FT $)+ Police use of smart glasses is also booming in China. (Gizmodo)+ LLMs could supercharge mass surveillance. (MIT Technology Review) 5 Space Force is awarding SpaceX $2 billion for a military data networkIt will connect military sensors and weapons platforms worldwide. (Reuters $)+ The contract comes amid concerns about SpaceX’s AI business. (WSJ $)+ Speculation is growing around a possible SpaceX-Tesla merger.  (CNBC) 6 Taiwan suspects Nvidia chips were smuggled to China via JapanTo circumvent US restrictions. (Bloomberg $)+ Is China about to win the AI race? (MIT Technology Review) 7 Booming AI chip demand has created two new $1 trillion companiesSouth Korea’s SK Hynix and the US’ Micron have hit the landmark. (BBC) 8 AI has sparked a surge in demand for cybersecurity expertsThanks to a glut of new code and alarm over powerful models. (NYT $)+ AI is making online swindles easier. (MIT Technology Review) 9 Internet is coming back in Iran after a three-month blackoutAlthough it isn’t clear if the reconnection is permanent. (Wired $) 10 Physicists are rethinking the role of gravity in quantum mechanicsThere’s a new theory for how our everyday world emerges. (New Scientist $) Quote of the day “AI and its capabilities represent something analogous to the Second Coming.”  —Jeremy Nixon, the cofounder of AGI House and a former Google Brain researcher, tells the New York Times how Silicon Valley’s innovations could affect the pope. One More Thing ANDREW MERRITT Inside the experimental world of animal infrastructure In the mid-2000s, toads were meeting a gruesome end near Ede, a leafy old town in the Netherlands. Residents responded by building wildlife tunnels beneath the road to help them reach their breeding ponds safely. The crossings became popular. But a few years later, researchers found the local toad population had crashed from more than 10,000 to fewer than 1,000. The case reflects a wider global push to build wildlife crossings and other forms of “animal infrastructure.” But do they actually help animal populations recover? Read the full story to find out. —Matthew Ponsford We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + The votes for “International Mollusc of the Year” are finally in.+ Track aircraft in real time across a gorgeous 3D digital globe using live flight data.+ NASA’s Psyche spacecraft has delivered breathtaking new close-up images of Mars.+ This deep dive into instant coffee reveals the extraordinary engineering effort behind making it vaguely drinkable.

In a

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. A reality check on the AI jobs hysteria Despite the growing hysteria over AI’s threat to white-collar jobs, there’s still scant evidence that the technology has had a large-scale impact on the labor market. Analysis of US labor data shows that unemployment in occupations most exposed to AI is actually lower than in less-exposed jobs. There are also no signs that large numbers of workers are shifting from AI-threatened professions into supposedly safer manual-labor jobs. It’s true that things aren’t great in the job market—but the question is why. Here’s what the data really says about AI and jobs. —David Rotman Opinion: It’s time to address the looming crisis in entry-level work —Georgios Petropoulos, an assistant professor at the USC Marshall School of Business AI has not yet produced mass unemployment. But it may be quietly weakening the first rung of the career ladder. A recent Stanford study found that young workers in AI-exposed occupations suffered a sharp decline in employment after the spread of generative AI. The same pattern didn’t appear in low-exposure jobs, suggesting AI is replacing junior tasks that once gave young workers their first foothold. It’s time to rethink how we train, prepare, and support young people entering the workforce. Read this op-ed on how job seekers, businesses, and society can adapt. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 The Pope has called for governments to regulate AI In his first major teaching document, Pope Leo said AI must be “disarmed.” (BBC)+ He warned that AI fuels war and misinformation. (CNN)+ But could also “open up a horizon extending in all directions.” (Engadget)+ Anthropic cofounder Chris Olah also spoke at the event. (Reuters $) 2 SpaceX has launched its biggest and most powerful rocketThe Starship V3 made its test flight debut two days after Elon Musk announced SpaceX’s IPO.(Guardian)+ SpaceX pulled off the launch, but not the landing. (Ars Technica)+ The rocket could be key to SpaceX’s valuation. (Fortune $)+ But rivals to the company are rising. (MIT Technology Review) 3 Huawei says it can make industry-leading chips within five yearsThe Chinese tech giant announced a breakthrough in chip design. (Reuters $)+ Its progress underscores Beijing’s push to neutralize US sanctions. (NBC)+ Chinese chip stocks rallied after the announcement. (Bloomberg $) 4 A new vaccine may protect against the Ebola strain behind the current crisisTests have shown promising results for the mRNA vaccine. (New Scientist)+ Another Ebola vaccine that could be ready for trials in months. (BBC)+ But vaccines face a new problem: their name. (MIT Technology Review) 5 A swimmer broke a world record at the ‘Steroid Olympics’Athletes at the Enhance Games were encouraged to take dope. (Wired $)+ Silicon Valley elites have backed the competition. (WP $)+ Which fits right into 2026’s longevity vibes. (MIT Technology Review) 6 The EU plans to fine Google a massive antitrust penaltyFor allegedly favoring its own services in search results. (CNBC)+ It would be the largest penalty for breaching the Digital Markets Act. (Reuters $)  7 US quantum computing subsidies may not be legalCongressional critics say the funding has been misused. (Ars Technica) 8 AI is minting new billionaires—and workers want their shareThe Samsung labor showdown reflects global concerns. (Rest of World) 9 China has launched artificial human embryos into orbitTo find out whether we can reproduce beyond Earth. (Gizmodo) 10 Jony Ives has designed Ferrari’s first fully-electric carThe legendary Apple designer has created a polarizing aesthetic. (FT $)  Quote of the day “Technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate, and use it.”  —Pope Leo issues a warning about AI in his first encyclical letter, entitled ‘Magnifica humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence.” One More Thing ALYSSA SCHUKAR How climate vulnerability and the digital divide are linked In Anacostia, a historic African-American section of Washington, DC, Monica Sanders is measuring Wi-Fi speeds. It’s below the FCC’s minimum to qualify as a broadband service. She then checks the temperature: 46.9 °F. Sanders, an adjunct professor of law at Georgetown University, frequently records this combination of weak internet access and environmental conditions. Her work shows how underinvestment in infrastructure can leave underserved communities more exposed to climate risks like extreme heat and flooding. Discover how the digital divide is shaping climate vulnerability in the US. —Colleen Hagerty We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Here’s a joyful way to settle sibling squabbles: a mandatory dance-off.+ Build the metropolis of your dreams in this browser-based city simulation game.+ Watch this hypnotic tiny train move in a perfect, endless loop on a rotating turntable.+ Take a nostalgic look at early computing history with this curated gallery of vintage punch cards.

Amid rapidly growing adoption of enterprise-level AI agents, there’s a disconnect emerging between ambition and execution.  Although 85% of organizations say they want to be agentic within the next three years, 76% say their current operations and infrastructure can’t support that change. They cite a lack of readiness across people, processes, and workflows.  The sticky tape problem The challenge is that many organisations are often layering AI agents onto existing operations, rather than reimagine the operating model and how work will need to be rewired, explains Prasun Shah, global CTO for workforce consulting and chief AI officer at PwC UK Consulting. “They’re embedding AI employees into what is a human operating model,” layering on AI agents to existing workplace structures when “this is like adding sticky tapes to parts of an operating model that is breaking.” Doing so may be preventing organizations from unlocking the full value agentic AI offers, creating circumstances where disillusionment can quickly creep in. That full value lies in agents’ capacity to execute entire workflows with limited human input. They can coordinate complex tasks, make independent decisions, adjust to changing conditions, and iterate performance.  In early proving grounds that span customer service, HR, and sales, it’s already estimated that AI agents could accelerate business processes by as much as 30% to 50% and low-value work time by 25% to 40% when deployed at scale. But with this capability comes greater complexity and the need for an enterprise-wide change. Growing the AI vocabulary  Enterprise agentic AI platform Ema describes this change as agentic business transformation (ABT), a term it coined last year in partnership with HFS Research, in an attempt to plug what it sees as a gap in the existing lexicon about AI agents, and to provide enterprises with a new framework with which to think about their own adoption of the technology.  “None of the existing vocabulary captures the full scope of the change,” explains Ema CEO and founder Surojit Chatterjee. “Digital transformation was about moving from paper to software. AI transformation was about adding artificial intelligence to existing processes. Co-pilot is about AI assisting in various human tasks. But ABT is something categorically different: It’s the integration of AI agents into the fabric of the organization.”  For Shah, the dedicated term (ABT) “helps drive the need to redesign an organization in its entirety: its operating model, its workflows, decision rights, and performance management systems.” He emphasizes that “everything that’s needed to ensure those agents are actually active participants in value creation, rather than just point tools or productivity aids.” According to Ema, ABT encompasses three core pillars: an organization’s technology stack, its workforce, and the metrics used for success.  AI agents as connective tissue The first pillar of ABT is the technology stack. “Your existing tech stack was designed for human-operated, application-centric workflows,” says Chatterjee. “It needs to be reconsidered when the actor is an AI agent operating at machine speed across multiple systems simultaneously.”  As AI agents are integrated into an organization, enterprises will need to pivot from a set of linear processes and steps, to rewiring work in a very different way, explains Shah. That’s because the value in AI agents isn’t as another layer in an existing technology stack but as a connective tissue, he explains, moving between or across layers to coordinate a high-level task or retrieve and interpret data from multiple discrete applications. AI agents can create “a true competitive differentiation for an enterprise” by making decisions based on this capacity to contextualize, he says. “That is where the next battleground will be.” To build this connective tissue, leaders need to adapt their technology stack to surface higher quality decisions from AI agents, prioritizing access to multiple datasets and applications simultaneously to develop tacit knowledge. “Organizations that make this architectural shift become genuinely more adaptive,” says Chatterjee. “When a new business requirement emerges, you don’t wait six months for a software vendor to build a feature. You configure an AI employee using natural language and connect it to the systems it needs. The time from business to production workflow drops from months to days.” The workforce, redesigned As AI agents are deployed for more use cases, enterprise leaders must consider what this means for dynamics across their workforce, the second pillar of ABT. Workforce structures today deviate little from the hierarchical model of the early days of industrialization. To maximize efficiency and scale, processes are standardized, tasks are clearly delineated between strategic business units (SBUs), and employees progress up through an organization based on their capacity to optimize output from teams below them. But with AI agents that can execute, coordinate, and optimize tasks—often without managerial coordination—the lines of that established hierarchy become blurred. In a workforce that blends AI agents and human employees, managers will be freed up from many execution-based tasks but take on new responsibilities associated with managing hybrid teams. Managers “will need to be able to manage issues around trust, explainability, psychological safety, and even status dynamics” to navigate new tensions that could arise in a hybrid workforce, says Shah. The impact of agentic AI on existing workforce structures goes far beyond the management layer, too. McKinsey predicts that by 2030, three-quarters of current jobs will require redesign, upskilling, or redeployment, and organizations will need to act swiftly to amend recruitment, retention, and remuneration.  From output to outcome Success metrics are the third and final pillar of ABT.  As AI agents assume greater ownership of core enterprise processes, taking on collaborative roles alongside human employees, traditional workforce metrics that focus on activity or output—such as calls handled or reports filed—no longer make sense.  “When you add AI employees into the workforce, activity metrics become meaningless or actively misleading,” says Chatterjee. “An AI employee can handle a thousand customer interactions in the time it takes a human to handle ten. If you measure success by interactions handled, you’ll conclude the AI is working brilliantly while missing whether any of those interactions actually drove customer satisfaction,