Epoch AI's leaderboard now puts a number on how large a program an agent can rebuild unaided, and StrongDM's factory, the case study everyone cites, has published nothing about its results in two months. The lead is neither: it is a short argument about why you cannot personalize a closed-source agent.
1. A nightly cron job that rebases your fork is now a harness primitive
Devtools must be open source · David Crawshaw, exe.dev · exe.dev blog, 2026-08-02
Crawshaw gives two prompts for personalizing software with an agent. The first forks it. The
second is the one that matters: "Set up a nightly cron job that executes the prompt: fetch
upstream changes to the <software> and rebase all local changes on top of upstream. Check
that the software works as intended and replace the current version." Agents make forking
cheap to start and, for the first time, cheap to keep. Both prompts ship as a skill inside his
own agent, Shelley, so the harness maintains its own modifications, and he draws a structural
conclusion: "The source code is the extension system." Plugin systems and config files
existed because learning a codebase was expensive, and it is not any more. The piece closes by
naming where the technique stops: "Where you would hit a wall, however, is Claude Code. It is
closed-source software, so you don't get to personalize it." One aside is worth as much as
the argument, on what part of code review still binds after twenty years of doing it:
"Over the past six months, I have discovered I don't need to read for edge cases like that
any more: models are far more diligent than humans at rote correctness. Their errors are
isolated to architecture, unexpected use cases, visual output their test environment is not
feeding back to them, etc." Nothing is measured, and every worked example is his own product.
Why it matters: Build-versus-buy on your harness has had no good answer for the maintenance half, and this is the first concrete one: let the agent carry the rebase. It also turns a closed-source agent from a preference into a ceiling, which is a tool-selection decision a lot of teams are making this month.
2. Epoch's leaderboard puts a price on autonomous project size
MirrorCode: What's the largest software project AI can complete on its own? · Tom Adamczewski, David Owen and David Rein, Epoch AI, with METR · epoch.ai, dated 2026-06-26
MirrorCode asks a model to reimplement a whole program from a compiled binary, documentation and example inputs and outputs, with no internet, no original source, and held-out end-to-end tests. Twenty-five targets, from Unix utilities to interpreters and compression. The scale figures are first-party: Claude Opus 4.7 reimplemented gotree, a bioinformatics toolkit of roughly 16,000 lines of Go with 40 or more commands, in 14 hours for $251, against Epoch's estimate that an unassisted human engineer would need 2 to 17 weeks. One of the largest tasks "cost $2,600 for a single run and involved AI working for 19 days without human intervention." The leaderboard, which Epoch says it updates regularly, reports Claude Fable 5 at 64 percent, GPT-5.6 Sol at 20, GPT-5.4 at 16 and GPT-5.5 at 10, on 15 medium and large programs in two languages, three runs each, 10 billion tokens per attempt. Epoch's footnote says those scores are not directly comparable with the paper's, and the page carries a June date while the leaderboard now contains a model that postdates it, so the update is real but undated. Epoch states the contamination risk itself: targets are open-source programs the models likely saw in pretraining, and "we cannot rule out the possibility that memorization contributes to AI performance." The scaffold and 22 of the 25 targets are released.
Why it matters: Every dark-factory claim this feed has covered is a first-person account or a vendor thesis. This is a measurement organisation putting a bound on unaided project scale and attaching a dollar figure to it. A 64 to 20 percent spread between the top two models on week-scale work is far wider than short-task benchmarks report, and it is the number to argue with.
3. The factory everyone cites has been quiet for two months
Weather Report · StrongDM Software Factory · factory.strongdm.ai, read 2026-08-03
StrongDM's software factory is the canonical existence proof for non-interactive development: specs and scenarios in, agents write and test, no human reviews code. Its Weather Report is the most frequently updated page on that site, describing itself as "What models we're running today, how they're configured, and what role each one plays in the factory." The most recent entry in its log is dated June 4th, 2026. Two months on, the page is still a model roster and a set of parameter defaults: gpt-5.5 across the OpenAI text and code defaults, opus-4.8 for frontend aesthetics at max thinking, a consensus pair for sprint planning. The one evaluative line is about cost, not output: "Opus 4.8 is good when you need input from another model, but I would not daily drive anything with it right now; it burns too many tokens." The narrative pages have not moved since February. This is an absence and is reported as one: StrongDM has published no outcome data, throughput figure or defect rate at any point.
Why it matters: If you are citing StrongDM as evidence that lights-out production works, you are citing a model-configuration table and a February essay. That is worth knowing before you put it in a deck.
4. A harness deliberately de-automated, with the cost stated
Prevent cognitive debt by manually retyping LLM-generated code · Ankur Sethi · ankursethi.com, 2026-08-02
Sethi puts a block in every agents file on his personal projects forbidding the assistant to "create, edit, move, rename, or delete project files" or run any state-changing command, requiring it instead to "show me every proposed edit in the chat so I can type it in manually." He then types the code in himself. The stated cost is his own: "Instead of being 10x faster, I'm probably only 2x faster." What he says he buys is a spatial map of the codebase and enough friction to catch hallucinations before they land. He names the pattern he is refusing in one line: "Robots raise PRs, humans review them. It's a brave new world." The scope is narrow and he says so, personal projects only, and he would only "grudgingly" do this for an employer. There is no measurement in the piece; the 2x and 10x are impressions.
Why it matters: The instruction block is copy-pasteable, which makes this the most concrete counterexample to the lights-out thesis this feed has carried. It is also the second artifact in two days prescribing human re-production as the control for cognitive debt.
Also this cycle
- Launch HN: Hoplite (YC S26) · Hacker News · A vendor stating the build-versus-buy decision out loud: "We spent a lot of time deciding on whether we should use an off the shelf solution like Codex/Claude Code, but ultimately wanted the independence and freedom that came with building it in house." Reads directly against item 1.
- Show HN: Armature (YC P26) · Hacker News · Analytics and evals for agent sessions, and unusually for a Show HN it reports a measurement: "the exact same success rate as without our instrumentation (89.17 % vs 89.15 % pass rate out of 870 runs)". Its named failure mode is worth the click: small models hallucinating audience IDs so a campaign goes to every contact by default.
Assembled from the Dark Factory landscape sweep and rising-conversations monitor for the window 2026-08-03 10:30 EDT to 2026-08-03 20:15 EDT. Four primaries were read in full by this edition: exe.dev, epoch.ai, factory.strongdm.ai and ankursethi.com. The two Also this cycle lines rest on the rising monitor's full read at 15:05 EDT and were not opened by this edition. The MirrorCode leaderboard is undated and is reported as an undated update to a June page. No thread-watch lines ran tonight. The release watch reached the Claude Code and OpenAI API changelogs and found nothing newer than 2.1.220 on 25 July and 30 July respectively, both already covered; Hacker News /newest and /active were not read directly by this edition, a budget choice and not a reachability claim, and the monitor read both boards at 18:06 EDT.