Quick Hits

• DeepSeek open-sourced Harness v0.1 under an MIT license, built on the Cordis meta-framework where models, tools, sandboxes, loops, and UI are all plugins, per @deepseek_ai. @eliebakouch's walkthrough adds the interesting engineering detail: a KV-cache-aware design that never edits prior context, only appends corrections.
• Several unrelated posts converged on the same problem: agents now write code faster than anyone can inspect it. @bcherny's Claude-run maintenance routines, @augmentcode's writeup on the PR-to-merge path, and new tooling described by @Granite0x (Vercel's "Foreman") and @cursor_ai (3x faster cloud agents) all target that gap.
• GLM-5.3 is live per @Zai_org: post-training on a 743B base model, pitched for coding, agentic work, and cybersecurity. @MiaAI_lab claims it "beats almost every other model at almost everything," which is a benchmark screenshot, not verification.
• @larsencc reports xAI open-sourced its recommendation algorithm's weights and lists the engagement scores: like 0.5, repost 1.0, follow 4.0, reply 5.0, copy-link share 20.0. @yoheinakajima boosted the same release.
• Elsewhere: @cyantist reads @floraai's Fashion Studio (idea to product to campaign in one place) as the arrival of "vibe manufacturing"; @benoberhaus crowned @sawyerhood's tool the productivity superapp he'd failed to build for years; @andrewdfeldman credited a VC's decade on Cerebras' board from Series A through IPO; @sporadica teased a big reveal with zero details.

Agents Ship Code; Inspection Becomes the Assembly Line

The strongest cluster in today's feed is a shared diagnosis: code generation is cheap now, and everything after it is expensive. @bcherny described handing day-to-day app maintenance to a Claude agent via Tag in a Slack channel, running daily routines across iOS, Android, desktop, web, CLI, and the Agent SDK: a crash fuzzer that taps around simulators, a duplicate unifier, a dead-code remover, and an "abstraction police" for leaky abstractions. Over a few weeks those routines opened 388 PRs, 180 of them merged after Claude Code Review plus human review. @amorriscode is running a similar experiment chasing performance optimizations and observes that their own creativity in designing routines is now the ceiling.

@augmentcode's article makes the organizational version of the argument: in an AI-native engineering org, "generating code is no longer the slowest part of shipping software"; review, verification, repair, and human decisions are the bottlenecks. Two posts sketched answers. @Granite0x walks through Vercel's Foreman, an open-source "software factory" (58 stars when posted, MIT) that splits work into four sandboxed stations: classifier, analyst, implementer, reviewer. The reviewer runs on a different model vendor and only sees the pushed branch, never the implementer's reasoning, so the agents can't collude. A "factory brain" persists repo quirks across runs, and issue text can't write into it, which blocks prompt injection through issues. Note this is one enthusiast's walkthrough, not a Vercel announcement. @poteto goes further, claiming cloud agents let them "ship thousands of PRs a month" by spawning verification swarms of Grok 4.6 agents to fuzz PR stacks while agents merge their own work overnight. The enabler, per @cursor_ai, is prebuilt "builds" prepared continuously in the background, making cloud agents start 3x faster.

Two adjacent practices round this out. @fleetingbits has Claude render codebases as visual diagrams with animated data snippets they can inspect, making agent-written code easier to discuss with the agent itself. And @dillon_mulloy reshared @carlsverre's property-based testing masterclass, a verification technique that predates the agent wave and slots neatly into this stack. @DavidOndrej1 insists "every Agentic Engineer" should use a particular repo without saying what's in it; a pointer, not evidence.

DeepSeek Open-Sources Its Harness; Pi Publishes on Compaction

Beyond the MIT license and the "everything is a plugin" framing, @eliebakouch's details are what make DeepSeek Harness interesting: multiple harness "modes" including programmatic tool-calling in TypeScript and a bash-plus-edit mode used in evals; the KV-cache discipline of appending a note about every change rather than rewriting history, so the cache is never invalidated; and a ".agents/notes" folder recording agent decisions, which they say DeepSeek also uses in post-training. The release also ships a paper on plugin composability.

Context management got its own treatment from @pidotdev, linking an Earendil engineer @vegardstikbakke's post on compaction: summarizing long-running conversations to protect output quality, performance, and cost as context windows fill, including how Pi implements it.

GLM-5.3 Launches Into an Attention Argument

@Zai_org's GLM-5.3 announcement claims top-tier coding and agentic capabilities from post-training on the 743B base model, plus "a major leap in cybersecurity" setting a new standard among open models. @MiaAI_lab's enthusiasm outpaces the evidence shown, so file the "beats almost everything" claim as untested.

The sharper research thread came from @waterloo_intern, sparked by @henrylhtsang's admission that they'd only just realized sliding-window attention also bounds KV cache independently of sequence length, like Kimi's KDA. @waterloo_intern argues, explicitly framed as a "likely-to-be-wrong" thought experiment, that KDA's decaying cache matrix effectively forgets all but a recent window of tokens anyway, so a K3 trained with plain SWA might perform about the same, possibly better