The agent loop went free this week.

npx @deepseek-ai/dsh web gets you DeepSeek's own agent harness: MIT, TypeScript, built on a plugin framework called Cordis in which models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI are all plugins. Four runtime modes, including a minimal one that is a two-tool coding agent with persistent bash and str_replace_editor, and a creator mode for building custom presets. The feature the thread actually liked is the append-only session log: everything the model sees is recorded, and you can resume, fork, search and replay it. SwellJoe's comment — "US models won't let you see that" — is the reason 674 points landed on it, though badlogic pointed out that Pi already shows and stores DeepSeek's CoT traces too.

Here is what verification turned up. The repo was created 2026-08-13 at 11:56 UTC, and the first commit visible is a merge of pull request #2519 against a tree containing a .gitlab-ci.yml and 221 packages, with 5,235 commits from one author — this is an internal repository published, not a project started. The npm package predates the repo: @deepseek-ai/dsh first went up 2026-08-10 at 19:41 UTC, the day before we wrote that the harness was unreleased. Six versions have shipped since, all release candidates through 0.1.0-rc.6; there are no GitHub releases and no tags. Issues are disabled on the repository, so the zero in the issue counter is a setting, not a state — discussions are on instead. And the benchmark disclosure, on the harness whose number was the reason anyone cared: BENCHMARK.md reads, in full, that you should follow the Python SDK quickstart and use separate workspaces and session IDs. No tasks, no trials, no build pinning, no cost, no result. The official page carries no figures either. DeepSeek gave away the machine and kept the measurement, which is the more valuable half.

Reach for it if you want a genuinely inspectable agent loop and can live on release candidates that say "THERE WILL BE COMPATIBILITY-BREAKING CHANGES" in the README. Delete the assumption that an open harness is a reproducible one. Tradeoffs: developer preview by the vendor's own label, with breaking changes promised rather than warned about; issues are off, so the only public defect channel is discussions; no tags or releases means no way to pin what you installed; the Cordis docs describe themselves as "A Meta-Framework of Spatiotemporal Composability," which one commenter reasonably read as impenetrable; nothing in the repo or on the site establishes which models it supports beyond DeepSeek's; the plugin-everything design drew the standard objection that community plugins are fine for six months and a nightmare afterward; and the 82.7% that made this interesting remains, three days and one full open-source release later, a number you still cannot reproduce.

NVIDIA's model router, Apache-2.0, Rust, pre-alpha. It sits between your client and your providers, speaks OpenAI Chat Completions, OpenAI Responses and Anthropic Messages natively in both directions, and routes each request to a configured backend by one of four strategies: an LLM classifier on content, a signal-driven stage router, a two-tier escalation router, and random for A/B. It deploys as a launcher in front of Claude Code or Codex, a standalone server, or an embedded Rust library, and it will talk to any OpenAI-compatible endpoint — vLLM, NIM, Ollama. No NVIDIA hardware required and NeMo is not required either, despite the org name.

The reason it is here, one slot below a repo with sixty times its stars, is the benchmark/ directory. It contains a Harbor Terminal-Bench Lite runner with a DATASETS.md that names the four datasets it prepares out of the box — TB Lite, Terminal-Bench 2.0, Terminal-Bench 2.1 and SWE-Bench Pro — a closed-book proxy with per-dataset allowlists so the agent cannot fetch its own answers, a run_manifest.py, a baseline shell script, and an agent-versions.env pinning CLAUDE_CODE_VERSION=2.1.211, CODEX_VERSION=0.144.5, OPENCODE_VERSION=1.18.3 and NODE_VERSION=20.11.1. Crucially it ships two paths against the same dataset: Harbor calling the provider directly with Switchyard disabled, and Harbor calling through Switchyard. That is a control condition, checked into the repository, on the same Harbor that Ante linked raw runs from on Monday. Be precise about what this is and is not: NVIDIA publishes the rig and no results. There is no results directory, no table, no committed run. So the two picks at the top of this edition are exact mirrors — DeepSeek has a number and no rig, NVIDIA has a rig and no number — and only one of those is fixable by a stranger with an OpenRouter key over a weekend.

Reach for it when you want to route across model tiers behind one endpoint and prove the routing is worth its overhead on your own tasks. Delete the vendor's comparison table, and the assumption that you need permission to run the experiment. Tradeoffs: it says "Experimental software. Not for production use." and means it — the 0.2.0 changelog removes the entire deprecated Python server stack, so anything built on the old path is already dead; 1,378 stars against 3 watchers and 92 open issues; three releases total since June 30; the Prometheus metrics cover routing overhead but no absolute latency or throughput figures are published anywhere; the benchmark path needs Docker with Compose and an OpenRouter or compatible key, so "cheap to reproduce" still means real tokens; and it crossed our pool on Tuesday inside a Nemotron launch we dropped as a model release, which is the third time this quarter something arrived as a launch and turned out to be a runtime.

A local desktop app that records agent sessions and links them to the commits they produced. It runs Claude Code, Codex and its own Rust agent as external subprocesses over ACP, and writes prompts, messages, tool calls and file patches into a SQLite database at .atlas/sessions.db, with a local embedding index and HNSW search over the history. The design decision that makes it work is that commits are observed rather than intercepted: a commit you make from a terminal while Atlas is closed still gets reconciled to its session, by patch-id, in a way the README says survives rebases and amends. MIT, 860 stars, trending at 249 a day, and the premise is stated plainly — "Agents now write a large share of the code and keep none of the reasoning behind it."

That premise is today's other front-page story stated as a product. Geoffrey Litt's essay took 348 points arguing that understanding, not verification, is now the bottleneck: "The agents are getting better and better at verifying their own work. And this is good!" — and yet the human stops being a participant. Atlas is the least glamorous possible response to that, which is why it is credible: it does not try to make you understand anything, it just refuses to throw away the transcript that would let you. It is also the only pick today whose repository contains a document I would hold up as a standard. TELEMETRY.md catalogues every event and property, names PostHog as the backend, points at the exact source files that emit, and then discloses a reversal against itself: "Through 0.2.3 this document promised that telemetry was anonymous permanently, and that signing in to an Atlas account would never be linked to it. That is no longer true, and the change was deliberate." A vendor writing down the moment it walked back its own privacy promise, unprompted, in the file where you would look for it, is rarer than any feature in this edition.

Reach for it if you run more than one coding agent and have ever tried to reconstruct why a three-week-old commit looks like that. Delete the habit of pasting an agent transcript into a PR description, and the guess about which agent wrote which hunk. Tradeoffs: macOS is the supported platform and the Linux and Windows builds are described as untested; the latest release is tagged alpha-0.2.6 and the tag list is mostly exp- builds; 175 forks against 10 watchers, and the contributor graph is one person with 283 commits and a long tail; it needs the claude CLI installed separately for Claude Code; usage analytics default to on, which the doc defends as the same opt-out posture as VS Code and Zed and which is still on; and the semantic index is another local copy of everything your agent read, on a disk you should now think of differently.

Codebuff relaunched its coding agent as a free one funded by advertising, took #3 on Product Hunt with 115 votes, and is trending on GitHub. You get a CLI, a desktop app, a web builder and a cloud agent, running DeepSeek V4 Pro as the default in full mode plus GPT-5.6 Luna, DeepSeek V4 Flash, MiniMax M3, MiMo 2.5, GLM 5.2 and Gemini 3.1 Flash Lite. npm install -g freebuff. Apache-2.0, 9,277 stars. The Product Hunt tagline is "Free coding agents to kill Claude, Cursor, Replit, and Devin," and the description names eight subscriptions you should cancel.

The sentence that matters is in the privacy terms, not the pitch: "We may analyze prompts and messages—including pasted content—to personalize ads." Uploads and connected repositories are carved out and stated as not provided to advertising providers, which is a real distinction and worth crediting. But the prompt is the product here — what you are typing to the agent is the targeting signal, and "including pasted content" is doing conspicuous work in a tool whose entire use case is pasting code and stack traces into a box. This belongs in today's slate because it is the price rung of the same ladder: yesterday's edition ended on the observation that nothing in the token-accounting category actually stops spending, and Freebuff's answer is to make the number zero and move the cost somewhere you cannot see on an invoice. It is an honest trade, clearly disclosed, and it is the first coding agent I have covered where the correct security question is not what the model does with your code but what the ad stack does with your prompt.

Reach for it if your budget is genuinely zero and your work is not confidential — students, side projects, throwaway prototypes. Delete the assumption that "free" and "no telemetry on your inputs" travel together. Tradeoffs: your prompts, including pasted content, are analyzed for ad personalization, and there is no paid ad-free tier described anywhere; the README makes none of the competitive claims the Product Hunt page makes, which is the fourth time this quarter PH copy has overstated what the repository supports; GitHub releases stopped at v1.0.420-beta.185 in October 2025 while the npm package has published 119 versions since March and sits at 0.0.149, so the versioning tells you nothing about maturity; 251 open issues against 48 watchers; limited regions get six one-hour sessions a day; model availability varies by "access tier, product, and current capacity," meaning the agent you benchmark today is not necessarily the one you get tomorrow; and nowhere does it publish a single quality number against any of the eight products it tells you to cancel.

Artificial Analysis, whose leaderboard has crossed our pool five times since June, shipped a product whose premise is that you should stop reading their leaderboard. Optima builds custom benchmarks: you describe your use case and import real examples or traces, an agent drafts candidate tasks and rubrics, you pick an evaluation type — Q&A, document input, agentic, or interaction — run it across the models you select, and get scores alongside cost and time per task. You can point it at your own agent over HTTP rather than at a raw model. Grading is either custom rubric criteria checked by a judge model, or their panel of judges from major evaluations ranking models head-to-head.

The pricing is the part worth reading twice, because it is the disclosure the rest of this edition is missing. Raw token costs pass through with "nothing added on top." Rubric grading is $0.125 per criterion, per model. Pairwise grading is $0.375 per match. Credits are held at the estimate and you are charged for actual use. That is a public unit price on the act of evaluation, which no one else in this slate offers at any price — and it prices the thing correctly, because the cost of an eval is not the tokens, it is the judging. Set it next to the top of the page: the most-starred artifact of the week is an evaluation harness that ships no evaluation, and the correction costs twelve and a half cents per criterion. The obvious objection is that a company selling model comparisons now also sells you the tool to disagree with its comparisons, and the judges in that tool are its own. It is a real conflict and the page does not address it. It is still the only offer on the board that ends with a number you did not have before.

Reach for it when a leaderboard rank and your actual workload have visibly diverged and you need to settle it before committing a model to production. Delete the assumption that a public benchmark describes your use case, and the internal spreadsheet where somebody eyeballs three models on five prompts. Tradeoffs: LLM-as-judge grades everything, including the agentic evaluations, and their own panel is the default — the same critique we applied to DeepTeam on August 10 applies here with the vendor's own thumb nearer the scale; the page states no limitations at all, which for an evaluation product is itself a finding; it does not say whether results are public or private, which matters enormously before you upload real traces; the "build agent" drafting your tasks and rubrics means a model is writing the test that another model sits; costs scale with criteria times models times tasks and the page gives no worked example of what a realistic run totals; and the Show HN drew 10 points and zero comments, so nobody has publicly run it yet.

Also in the pool

mcp-stama took 71 points for a single-binary Rust MCP server with no dependencies and has zero comments and 5 stars on a repo created yesterday — a points-to-traction gap wide enough to be worth naming rather than covering. MCP Memory (129 stars, MIT, built yesterday) got the better thread and the harder question: bearjaws called it "another agent memory system that is about the same as grep in a memory/ directory," and clemens1010 asked whether the author had "test[ed] if that actually outperforms local claude code memory by any metric." He had not, which is the edition's thesis arriving from a fifth direction. NanoRL is an async RL trainer in about 1,800 lines across seven files with no Ray, TRL or DeepSpeed, running CartPole on a laptop and GRPO on 8×H100 — MIT, 5 stars, meant to be forked rather than imported. Lumabri re-crossed after landing on August 10 as "What if LLMs worked like Napster?", now framed as MoE inference over a P2P swarm in pure C, 71 stars. Dropped per rubric: Gemini 3.7 Flash (867), Cerebras on GPT-5.6 Sol (621), GLM-5.3 (540) and Mistral OCR 4.1 (363), all launches. Also dropped and worth your time anyway: Pi's writeup of how compaction works, which specifies a 20,000-token retention budget of roughly five to twenty turns, stores its summary as plain text you can read, and admits the cost — compaction breaks the prompt cache, every time. Fifth Pi crossing since May, and still the only harness in this edition that has published how its context loop actually behaves.

Pool: 78 candidates (HN front 30, ProductHunt 16, HN Show 15, GitHub trending 17 across five feeds). Reddit returned zero for the fifteenth consecutive week.