Leitura

As enterprise AI agent platforms become more prevalent, employees find it easier to build and deploy agents to support their own work. However, this decentralized creation introduces a new management challenge: how to track what agents have been built across the organization. In a blog post published on October 15, 2026, Cohere presents a counter-intuitive view: the increase in the number of agents (proliferation) does not necessarily lead to chaos. It should only be termed 'agent sprawl' when the process for managing agents fails to keep pace with their growing numbers.

The article first distinguishes between 'proliferation' and 'sprawl'. Those closest to a workflow are often best placed to determine how it could be streamlined by an agent, making decentralized experimentation valuable. Furthermore, an agent that isn't being used doesn't consume tokens simply because it exists, meaning pure quantity growth doesn't directly equate to high operational costs. The real pain point lies in a lack of visibility, which makes it difficult for teams to find and reuse existing agents and harder to govern access to enterprise data and systems.

Currently, enterprises typically manage growing agent estates through centralized, federated, or hybrid delivery models, increasingly using agent registries to maintain a shared view. These methods share a common preference for avoiding duplication, encouraging teams to default to reusing existing solutions. While reducing overlapping options lowers monitoring and governance burdens, Cohere argues that treating duplication as an absolute taboo can suppress innovation. If the mere existence of a workable agent means no alternatives should be built, organizations forfeit the opportunity to discover solutions that are cheaper, more reliable, or better suited to the task.

To address this, Cohere proposes an alternative approach: allow agent variants to compete. In this model, duplication is treated as provisional rather than permanent. If a team has a credible reason to believe it can improve on an 'incumbent' agent, it should be permitted to build a 'challenger' variant. However, the continued use of the challenger depends on whether it demonstrates a meaningful advantage over the incumbent.

This competitive model requires a time-bounded, evidence-based comparison. Specific metrics will depend on the use case but could include measures such as reliability, token use, latency, and task performance. For example, consider a bank's commercial lending division where a central team has built a finance agent that drafts credit memos. A lending team handling routine annual reviews might build a challenger designed for the same tasks using a smaller, cheaper model and an additional cross-checking step. For 30 days, both agents draft memos for representative cases, and the bank compares them on four measures: how many figures trace correctly to source documents, how heavily credit officers edit each draft, token cost per memo, and time to first draft.

Once enough evidence is collected, the organization decides whether the challenger should replace the incumbent or be retired. If the challenger proves better suited to a part of the workload, both agents might remain as distinct alternatives. The key is to pair variation with selection pressure. By making agent creation easy but persistence conditional on performance evidence, enterprises can preserve the benefits of continuous experimentation without sliding into uncontrolled agent sprawl.

When building the new standalone Groups app Forum, Meta's engineering team experimented with treating AI agents as participants in the engineering process, rather than just faster code-completion tools. The core shift was moving from 'human-in-the-loop' to 'human-on-the-loop'. In the old mode, engineers drove every step with AI assisting; in the new mode, agents autonomously run iterations of planning, implementation, building, testing, and fixing, while engineers set goals, observe runs, and intervene when deviations occur. This transition required defining success not by lines of code, but by verified, production-ready results.

To enable agents to run safely and autonomously, the team established a standard where verification, not compilation, is the endpoint. A task is only complete when automated tests and visual checks pass. To support this, they implemented four key capabilities: first, treating planning as technical specifications with machine-readable context to reduce ambiguity; second, integrating automated pass/fail signals like build status, unit tests, and UI spec matching directly into the agent loop; third, deploying navigation agents for simulated user testing that interact with live apps like humans and report regressions; and finally, using agentic production operations to assist with root cause analysis post-launch.

The biggest gains came from adapting the environment to the models, not just optimizing prompts. The team prioritized architectural clarity over inherited complexity, choosing modularity so agents could easily plug in. Crucially, they gave agents tools to verify state rather than reasoning based on descriptions. For example, agents can read real-time view hierarchies from simulators, list screen elements, or find paths to target views, significantly reducing erroneous edits. Additionally, independent build variants provided sandbox environments where prototype code was completely excluded from production builds, allowing fast AI iteration while protecting stability.

This workflow uncovered real defects. During routine stress testing, an agent exposed a memory issue, automatically isolating reproduction steps, proposing a fix, and verifying it. The fix significantly reduced device memory growth and page faults. Forum reached core feature parity with Facebook Groups in eight weeks, with AI handling most routine implementation work, freeing engineers to focus on architecture and product decisions. However, this method relies heavily on robust automated validation infrastructure, which may present a high barrier to entry for teams lacking modular architectures or comprehensive test coverage.

In the development of large language models, the training costs of dense models rise sharply as parameter counts increase, making simple model scaling unsustainable due to hardware resource limitations. Mixture-of-Experts (MoE) models, as a sparse architecture, can significantly reduce computational costs during training; however, their massive parameter count leads to a surge in memory requirements and low inference efficiency, limiting their practical application. The DeepSpeed-MoE paper addresses this pain point by proposing an end-to-end solution aimed at balancing model quality, training cost, and inference efficiency.

One of the core contributions of this research is expanding the scope of MoE applications and proposing new architectural designs. Traditional MoE models are mostly used for encoder-decoder tasks, but this paper successfully applies them to autoregressive natural language generation (NLG) tasks, demonstrating that a 5x reduction in training cost can be achieved while maintaining equivalent model quality. To address the issue of low "parameter efficiency" in MoE models, the authors designed Pyramid-Residual MoE (PR-MoE). This is a hybrid dense-sparse architecture that applies sparsity only where experts are most effective through residual connections, thereby reducing the parameter size of MoE models by up to 3x without changing model quality or significantly altering computational requirements. Furthermore, using staged knowledge distillation techniques, the authors created a distilled version of PR-MoE called Mixture-of-Students (MoS), which further compresses the model size to 1/3.7 of the original while maintaining comparable model quality.

Beyond architectural innovations, DeepSpeed-MoE also developed a highly optimized inference system to solve deployment challenges for large-scale MoE models. Since MoE models are much larger than equivalent dense models, inference processes are often limited by memory bandwidth rather than compute capacity. By combining model compression with system-level optimizations, this system achieves significant inference acceleration and cost reduction. Specifically, compared to existing MoE inference solutions, DeepSpeed-MoE provides up to 7.3x improvements in latency and cost; compared to dense models of equivalent quality, its inference speed is improved by 4.5x and costs are reduced by 9x. This enables ultra-low-latency inference below 25 milliseconds for trillion-parameter-level MoE models, providing a feasible engineering path for transitioning from dense models to sparse MoE models.

For readers interested in infrastructure optimization for large models, DeepSpeed-MoE offers a clear perspective on how to break through compute bottlenecks via hardware-software co-design. It is not merely an algorithmic improvement but a complete system solution, demonstrating how to deploy high-quality large models under resource constraints. For example, if a team needs to deploy a high-performance dialogue model on a limited GPU cluster, directly adopting a dense model might fail due to insufficient VRAM or excessive inference latency. However, by referencing the PR-MoE architecture and inference optimization strategies in DeepSpeed-MoE, they could potentially maintain response speed while significantly lowering hardware barriers.

Engineers and researchers engaged in AI systems engineering, model compression, and efficient inference research are recommended to read this paper, "DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale" (2022), focusing particularly on Section 1 regarding the description of challenges and Section 3 regarding technical details of PR-MoE and the inference system.

In his 1987 letter to shareholders, Warren Buffett delves into the distinction between book value and business value, articulating Berkshire Hathaway's core investment philosophy. First, he explicitly states: "What counts, of course, is the rate of gain in per-share business value, not book value." To illustrate this point, he cites examples such as LTV and Baldwin-United, which displayed high book values prior to their bankruptcies, and Belridge Oil, which was sold to Shell at a price far exceeding its book value. This serves as a reminder to investors that accounting figures often fail to reflect a company's true economic condition; real value lies in the enterprise's ability to generate cash and its competitive moat.

Second, Buffett emphasizes the value of operational stability, opposing the blind pursuit of so-called "high-growth stories." He writes: "Experience, however, indicates that the best business returns are usually achieved by companies that are doing something quite similar today to what they were doing five or ten years ago." He argues that while frequent major changes may sound appealing, they also create opportunities for significant errors and make it difficult to establish durable business franchises. Citing research from Fortune magazine, he notes that only a few companies maintain extremely high return on equity over a decade, and most of these operate in seemingly mundane businesses—products or services that are not glamorous but are executed exceptionally well.

Finally, the letter showcases Berkshire's unique decentralized management system. Buffett mentions: "With managers like ours, my partner, Charlie Munger, and I have little to do with operations. In fact, it is probably fair to say that if we did more, less would be accomplished." This trust stems from identifying excellent managers, such as those leading subsidiaries like Buffalo News and See's Candies. They not only generated astonishing capital returns but also allowed headquarters to focus on capital allocation rather than micromanagement. For today's readers, this provides a clear framework: when evaluating a company, one should focus on the durability of its business model and the autonomy of its management, rather than short-term financial embellishments or aggressive strategic pivots. If you are analyzing a traditional consumer goods company whose product lines and profit models have remained stable over the past decade, and whose management enjoys high autonomy, it may possess greater long-term investment value than a tech company constantly launching new concepts but lacking a track record of execution.

For investors seeking to understand how to see through financial statements to grasp the essence of a business, and how to construct a decentralized portfolio driven by outstanding managers, the Berkshire Hathaway Shareholder Letter 1987 is a must-read classic. It clearly defines the difference between business value and book value, and argues for an investment logic where stability outweighs change. It is recommended to start reading from the opening paragraphs regarding the definition of value, then focus on the description of subsidiary operating models.