Amarnath Byakod’s article, “Context Engineering and Knowledge Cataloging for AI Agents,” begins with a problem that many enterprise AI teams will recognize. Documents are divided into chunks, converted into embeddings and stored in a vector database. An agent retrieves passages that appear relevant and produces a fluent answer. The demonstration looks successful until someone asks where each statement came from, whether the passages belong together and whether the answer reflects the correct business context. — Read More

#devops

… Programming with LLMs is genuinely useful and genuinely destabilizing. These two things coexist. If we pretend the second one isn’t happening, we will all burn out.

…[I]n the last few months, there have been days when I have spent close to two full days writing a plan for an LLM to execute: obsessively clarifying, specifying, re-specifying, only to have it still do something inexplicably stupid. Port a React hook into a Storybook story file. Read from the wrong plan. Invent components that don’t exist. And these aren’t errors of capability; they’re errors of coherence. The models are smart enough to produce plausible code, but not always smart enough to maintain a coherent intent across a complex change.

This creates a peculiar new kind of fatigue, the fatigue of supervision: of holding the intent in your head while the machine generates volumes of mostly-correct output that still needs your eyes, your judgment, and your taste. Douwe put it well: he used to get a dopamine hit from collaborating with a real person on a cool feature in open source. Helping someone become better at their craft. Now, he said, “everything I write goes into some AI black hole. There’s no person on the other side actually learning anything.” That loss is real and it’s worth naming: the intensity trap. — Read More

Embedding complex workflows and logic into AI Agents through skills alone is a dramatically inefficient way to work. By using a swamp workflow instead of the skill alone, I dropped the token usage 8x, and decreased the run time by 2x for a complex code review workload. For a lot of use cases, it’s the single best thing you can do to reduce your token spend. — Read More

#devops

Open an open-source coding-agent codebase and you may want to see how it actually runs, verify that it behaves as safely as the documentation claims, or adapt it into an agent of your own. These goals sound different, but once you are in the code they all reduce to concrete questions about behavior. Will the agent ask before deleting a file, for example? Answering that means finding the confirmation logic, tracing bypass paths, and identifying every implementation site a change would touch. In a repository with thousands of files, searches for delete, permission, or confirm return scattered fragments—and piecing them back into a full behavior chain is hard work.

The problem is not missing code, but missing a path from behavior to implementation. What is needed is not another code index, but a map that connects the two. Harness Handbook organizes scattered implementation into a behavior-level manual: it structures execution around system behavior and links every step to verifiable code evidence. Users can ask directly what they want to understand, audit, or change, and the Handbook locates the relevant behavior units, implementation sites, and next steps. As the harness evolves, this map keeps the system understandable and reviewable—and keeps humans in the loop throughout. — Read More

While a lot of focus goes to advances in large language models (LLMs), the performance of an AI application is largely dictated by its runtime harness: the execution logic, system prompts, memory management, and tool configurations that connect a model to the real world.

Developers want custom behavior from their applications, but training a model from scratch or fine-tuning open-weight LLMs is too expensive and fraught with different complexities. For most engineers, the harness is the most accessible lever for control.

As new models drop rapidly, manually updating and crafting these harnesses per model scales poorly. Harness optimization has remained a manual, time-consuming chore.

Recent AI frameworks are reframing this constraint. Instead of relying on manual labor, these frameworks structure the harness so that AI agents can iteratively analyze, test, and optimize their own runtime environments. — Read More

Autoresearching the autoresearch agent for eight days.

The result beats the harness we hand-tuned for two years, on held-out benchmarks

Our RSI system AIDE² has two autoresearch loops.

An inner loop, just like a normal autoresearch agent, optimizing code against an eval.

An outer loop, optimizing the inner-loop agent’s harness code against the inner loop’s average score across different benchmarks. — Read More

AI agents perform real work on our behalf: editing files, running code, and carrying out multi-step tasks over hours or days. This often requires access to private information, which makes where an agent runs matter more than how smart the model is.

Sandboxes keep agents secure, letting them access the tools and files they need to complete these tasks, while keeping sensitive information protected. 

When building our agent platform Computer, we explored off-the-shelf sandbox options. Every option required us to consider tradeoffs between functionality, efficiency, and security. Perplexity Computer needed all three, so we built an entirely new type of sandbox.

Today, we’re introducing SPACE, a sandbox platform designed to unlock Computer’s full power while providing the highest level of security for advanced agentic systems. — Read More

For the last three years, the default engineering response to any enterprise AI context problem was automated: “Just build a RAG pipeline.”

… But as we advance through 2026, the cracks in the RAG-everything approach have become too wide to ignore. Chunking destroys complex table structures, vector retrieval is inherently probabilistic (you might get the right chunk, or you might get an outdated one), and keeping embeddings synchronized with rapidly updating data is an absolute operational nightmare.

To solve this, Google Cloud quieted the “RAG-everything” noise by open-sourcing the Open Knowledge Format (OKF v0.1). It isn’t a new cloud database, an LLM framework, or an SDK. Instead, it is a vendor-neutral, portable specification that formalizes the “LLM Wiki” paradigm — the exact structured, interconnected “brain” concept long advocated by AI researchers like Andrej Karpathy. — Read More

Most people building with AI are one or two layers deep in a seven-layer system.

Every week someone asks me what tools they should use to build their AI system. The answer is always the same: it depends which layer of the stack you are talking about — and most people asking the question have not The Full AI Stack Explained: Every Layer, Every Tool, and How to Make the Right Choice at Each One — Read More

I’ve been asked on multiple occasions to produce a blog post covering programming with LLMs from A-Z. In this post I’ll discuss all the most important terms and definitions as well as a multitude of DOs and DONT’s with LLMs and writing code.

These statements will of course be my own opinions, however I hope you’ll be agreeing with all of them by the end 🙂 — Read More