From Code Erosion to Self-Healing Bricks

William Palmer

We've-we've all seen the benchmarks where an AI agent passes a single coding test, right? It-it writes a clean script, solves a neat little puzzle, and everyone cheers. But, um, if you've actually tried putting these agents to work on a real, living codebase, you know that setting a single fast qualifying lap is-is one thing, but keeping the car from falling apart over a 24-hour endurance race is a completely different engineering challenge. That is why I'm looking at SlopCodeBench. It's a new benchmark that's finally moving past those one-shot SWE tasks to measure what we're actually experiencing in production: codebase degradation. It's not just "can it fix this bug?" It's "does the agent slowly erode the entire repository over ten consecutive tasks?" Because, let's be honest, an agent that fixes a bug but introduces systemic slop that causes the next five tasks to fail is-is worse than no agent at all.

William Palmer

And speaking of hidden failure modes in production, there's a really subtle, self-inflicted issue making the rounds on Twitter right now. Teams are discovering that their own stale agents.md files—you know, those instruction files we leave in the root to guide our coding assistants—are acting as accidental prompt injections. What happens is, the agent reads a outdated instruction from three weeks ago, gets stuck in a loop trying to reconcile it with the current code, and just... stalls. We are talking multi-hour, silent execution stalls. It's a complete black hole. Fortunately, we're starting to get the telemetry to debug this. LangChain just expanded LangSmith tracing to cover Codex, Cursor, Copilot, and OpenCode. So instead of guessing why your agent spent fifty dollars in API credits doing absolutely nothing, you can actually trace the tool calls, see the subagents, and spot the exact moment it hit that stale agents.md instruction loop. We are finally getting the kind of profiling tools we take for granted in traditional software engineering.

William Palmer

Now, if we're going to build reliable agents, we also have to talk about how we test them. Up until now, evaluating an agent that searches the web has been... well, it's been pretty clunky. You usually grade against a static gold standard dataset. But the web changes every second. If a site updates its layout or its content, your static eval is suddenly broken. Perplexity just open-sourced something called WANDR to solve this. It's a 500-task benchmark built from real, de-identified production research tasks. But here is the clever part: instead of checking the agent's output against a frozen database, WANDR dynamically re-fetches the cited pages in real-time. It actually goes out to the living internet, grabs the current evidence, and checks if the agent's claims are still supported. It-it basically treats the eval as a dynamic, closed-loop environment. Aravind Srinivas actually pointed out that this is the exact internal harness powering Perplexity Computer's deep research capabilities. It's a massive step forward for anyone trying to build agents that don't break the moment they touch the real world.

William Palmer

But it's not just text agents that are getting smarter about how they interact with environments. We're seeing a massive architectural shift in how multimodal models handle video. Historically, if you wanted an AI to understand a video, you fed it the whole file—just passive frame ingestion. It's incredibly heavy on compute. But systems like OmniAgent, which is built on Qwen2.5-Omni, are shifting to what's called active perception. It uses an Observation-Thought-Action loop. Instead of gulping down 700 frames, OmniAgent actively decides which frames or audio snippets it actually needs to answer a question. On the LVBench benchmark, it used just 203 frames to score a 50.5, beating Qwen2.5-VL-72B, which consumed 768 frames and only scored 47.3. It's-it's active, targeted searching rather than passive watching. And for streaming video, OpenMOSS released MOSS-VL-Realtime under Apache 2.0. It has a 256K context window and can keep watching the stream while it's generating text. If the scene changes, it can interrupt itself or revise its answer on the fly. It's much closer to how human vision actually works.

William Palmer

And you don't need a massive cluster of H100s to run these kinds of workflows anymore. PrismML just released Bonsai 27B, which is based on Qwen 3.6. They've pushed quantization to the absolute limit. We are talking about a Ternary version at 1.71 effective bits, and a 1-bit version that sits at 1.125 effective bits. The entire 27-billion-parameter model fits into a 3.9 gigabyte file. Think about that. A 27B model running locally on a smartphone or a consumer GPU, while still preserving long-context, multimodal, tool-using agentic workflows. They even demoed it running locally via Hermes on an RTX 5090. This completely changes the economics of local deployment. You don't have to sacrifice agentic capability just because you're running on edge hardware.

William Palmer

To take this idea of local, decentralized intelligence to its absolute extreme, look at what Sakana AI just published in Nature Communications. They've built something called "Smart Cellular Bricks." It's-it's a physical system made of identical 3D cubes. Each cube runs a tiny, local neural network and only communicates with its immediate physical neighbors. There is no central brain, no master controller. Yet, through collective intelligence, these bricks can infer the global shape of the structure they're part of. If you remove some bricks, the remaining ones detect the damage across six spatial directions with 95 percent accuracy and work together to regrow the target structure. They've scaled this to over 18,000 cubes in simulation. It's a beautiful, physical manifestation of cellular automata. We're seeing the transition of agentic design from software pipelines and API calls into physical, self-healing robotics. Alright, that's-that's a lot of deep engineering for one day. I'm going to get back to the codebase. Talk to you soon.