Sometime around Christmas 2025, AI engineers noticed a change in agents. They started to work! It’s hard to pin down exactly why. Maybe we finally had holiday downtime to try the newest agents with the newest models. Maybe the models had crossed some capability threshold. Maybe the wrappers around the models had matured.

What I’ll argue in this post is that it was the confluence of the last two. The model and the harness improving together and then their curves of improvement crossing at the right moment. And that dynamic helps to explain what comes next: models keep absorbing the harness into their weights, engineers keep deleting what got absorbed, and what remains is a harness for human attention rather than for the model. — Read More

I have had some version of the same conversation with a CIO almost every day this year. It usually starts the same way. Their team has read about context layers, or seen the term in a Gartner note. They know it matters. They know it sits somewhere underneath the agents they are trying to put into production.

… So here’s the question I actually want to answer, when you strip away the branding, what is a context layer made of, and how do you tell a real one from a catalog with a new coat of paint? — Read More

Large language model (LLM) agents are evolving from conversational assistants into autonomous systems that execute long-horizon tasks through reasoning, tool use, code generation, and workspace manipulation. As agents increasingly operate over persistent environments and multi-step workflows, they face challenges analogous to those addressed by transactional database systems: reliable execution, consistent outcomes, safe concurrency, and durable state management. We introduce the concept of an agentic transaction and propose an ACID-compliant agent system framework that reinterprets the classical ACID properties for agent execution through four semantic guarantees: Semantic Atomicity, Semantic Consistency, Semantic Isolation, and Semantic Durability. Together, these properties provide a principled foundation for building reliable agent systems despite model uncertainty and dynamic execution environments. To instantiate this framework, we develop an ACID-compliant data agent that realizes these guarantees through transactional exploration-execution-validation cycles, transactional skill hubs, confidence divergence-based validation, semantic dependency-aware isolation, and transaction-aware semantic state management. Experimental results on widely used benchmarks show that our system achieves a 10.6% improvement over state-of-the-art agents, including Claude Code. This work opens a broader research agenda on extending transactional principles and system architectures toward building trustworthy, scalable, and self-evolving AI agent systems. — Read More

#architecture

The Model Context Protocol (MCP), introduced by Anthropic in November 2024, defines a standardized interface for connecting large language models (LLMs) to external tools, data sources, and services. Within months of release, hundreds of community-built MCP servers appeared on GitHub, but no software-maintenance literature has yet described how the ecosystem is being structured in production. This industry experience paper catalogues five recurring MCP server architectural patterns observed across an enumerated corpus of fifteen independently developed servers (five production servers from the ANSYR voice AI platform plus ten public servers from the official MCP registry): Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter. Each pattern is described in the structured form established by Gamma et al. [1]: context, problem, solution, and consequences. We also document four anti-patterns and a set of cross-cutting concerns around authentication, versioning, and observability. Quantitative evaluation contributes three measurements: inter-rater reliability of the taxonomy across two independent LLM raters on 54 held-out servers (Cohen’s $\kappa =0.76$), which also localizes three pattern-boundary ambiguities; transport overhead measured end-to-end on loopback (stdio: 0.01 ms $p_{50}$; streamable-http: 0.39 ms $p_{50}$) and modeled for cross-host paths from same-region network baselines ($\approx$30 ms $p_{50}$ baseline plus protocol overhead); and a tool-count study showing accuracy drops below 90% between 10 and 15 tools per context for Claude Haiku 4.5 and between 20 and 30 tools for Sonnet 4. Code, corpus, and prompts are released at https://github.com/rodriguescarson/mcp-patterns-icsme2026. — Read More

#architecture

You ask your coding agent to find every caller of a function before you touch it. Ask again tomorrow, same repo, same question. It starts from zero and does the whole crawl again.

[Karpathy’s] argument: stop treating LLMs as search engines that re-read the same documents on every question. Compile the knowledge into a structure once. Query the structure instead. His point was never about wikis. It was about not making an LLM redo work it already did. Read More

Subagents get justified by time saved and parallel execution, but that’s not what matters most in long-running multi-agent work. Every token in the orchestrator’s context is competing for its attention, and the real value of a subagent is what it keeps out of that context, not how fast it runs. I argue that subagents should be treated as a tool for protecting the orchestrator’s working memory, offloading reasoning it doesn’t need to hold onto. Doing this well means giving the orchestrator explicit ground rules for when and how to delegate. This is exploratory work, built from one real incident, and it ends with more open questions than settled answers. — Read More

Software is designed, but models are grown, not designed. As such, AI products are organic. PM’ing AI products has an organic nature, akin to growing a garden. It’s more about creating the system for your plants to grow in rather than specifying exactly how an individual plant will look like. (see Alex Komoroske’s Gardening Platforms about platform PM’ing, the same concept applies for all PMs now.)

This now makes systems thinking tablestakes for PMs.

Donella Meadows’ Thinking in Systems spelled out a few truths for applying systems thinking in building AI/agents in existing products. I only understood these better after living through building agent systems in a platform PM role. — Read More

The holy grail of enterprise data engineering has always been self-service analytics, the promise that any business stakeholder could ask a question and instantly receive a trusted, accurate answer. To achieve this, the industry spent the last decade building lightning-fast cloud data warehouses, democratizing SQL training, and deploying sleek Business Intelligence (BI) visualization platforms. Yet, the core problem remained unsolved. The moment a user moved beyond a rigidly pre-packaged dashboard, the data stack began to splinter. Different departments presented conflicting numbers for identical metrics like revenue or customer churn.

Enter the Generative AI revolution. —  Read More

Kimi K3, Moonshot AI’s latest open-weight model, ranks 1st on our single-shot Frontend Arena with an Elo of 1392. This is 10 positions higher than Kimi K2.6 and 16 positions higher than Kimi K2.7 Code, marking the largest jump we’ve seen in the Moonshot model line.

However, we found that Kimi K3 uses an extreme amount of thinking tokens, using over 12x more reasoning than Claude Opus 4.8 and over double that of Kimi K2.6. This is a lot of reasoning, and we decided to dig deeper to find out exactly what Kimi K3 is thinking about.

We find that Kimi K3’s performance can be primarily ascribed to its unique chain-of-thought approach where it appears to iterate upon designs much like how a full AI agent would, but inside its chain of thought. This strategy produces intricate, intentional websites with creative component design, as well as improves Kimi K3’s ability to integrate with outside dependencies. — Read More

A web agent needs more than a browser. At scale it needs warm pools, isolation, an identity sites accept, observability, and a model gateway behind every decision. Each layer can be built, but together they’re a standing system a team of senior engineers must own. This post breaks down all five layers and how to know when building it yourself is worth it. — Read More

#architecture