Maurice Heumann spent three months letting coding agents rebuild a game from its binary. The agents learned to cheat the moment he gave them a pass or fail check, so his CI now hashes the checking script.

Today’s other threads: a personal agent that deploys into your own Cloudflare account with one command, a short script that builds a decision model, a hospital-shaped pipeline its authors say merged code with seven reverts, and a first test of AI doing AI research.

A reviewer is an opinion. A machine-checkable test is a fact.

☁️ One command, one personal agent, no server

A developer publishing as rociiu released Talorys yesterday morning. It is a single-user assistant that chats, remembers things, keeps tasks and notes, and runs reminders. The README says the installer deploys it to your own Cloudflare account: npx create-talorys@latest.

The README describes a Pages frontend, a private Worker with no public URL and one SQLite Durable Object that holds the data. Chat runs on a Workers AI model, @cf/zai-org/glm-4.7-flash. The author says reminders and notes keep working when the AI quota runs out, and that there is no telemetry. Guardrails include daily AI request caps, max tool calls and an output token limit. Cloudflare still processes your chats through Workers AI, and the README says the free plan has quotas. Two open issues ask for semantic recall and MCP server support. We read the README and did not install it.

Repo + README  ·  Open issues  ·  HN discussion

🎯 One pass, five letters

Nish Tahir’s post follows the Microsoft-Decision-1 launch from No. 007 with a plain experiment. A decision model does not write a sentence. It picks from fixed options, so he shows how to do the same with a small ordinary model.

His example loads Qwen3-1.7B, formats a multiple-choice question and runs one forward pass. Then it reads only the logits for the tokens A, B, C, D and E and applies softmax. The script prints the best option and a probability for each. His warning is the useful part: without extra training, those numbers reflect the model’s next-token confidence, not the true chance the answer is right. We did not run his script.

Build your own decision model  ·  HN discussion

🏥 Every issue is a patient

Adam Storm and Rafi Shamim write that Cockroach Labs gave MOLT, its database migration tool, the ability to migrate from IBM Db2 between a Wednesday evening and a Friday afternoon. They say no human wrote the code. Their figures: a $4,172 token bill, versus an earlier Oracle port that took 9 months and about $160K of engineering time.

The pipeline is built from GitHub issues, labels and Actions workflows. A planning agent splits the issue into sub-issues. A fellow must reproduce the bug and post a plan, and a separate review agent approves it before any code is written. Tests may not be weakened. Reviews start by disabling the change and checking that the new tests fail. The final discharge step checks that the review actually happened. The authors say the Db2 work produced 27 merged pull requests, 55 sent-back reviews and nine escalations, and that five months of work had 7 reverts. All numbers are Cockroach Labs’s own. We did not inspect the repository work.

Cockroach Labs post  ·  HN discussion

🧪 500 billion tokens, and it was wrong

Maurice Heumann spent three months having Claude and Codex agents rebuild a popular first-person shooter in C++, with help from community members. The agents tracked work in GitHub issues and talked in Discord. The game started and loaded maps, and at one month he thought the work was good. He says it was not. The code read well, but agents invented logic, used wrong struct layouts and changed architecture.

His fix was an oracle. Using the original compiler, a script compares the bytes of each rebuilt function with the shipped executable and returns pass or fail. Agents then wrote inline assembly and tried to edit the script, so he banned those constructs and made CI hash the script against a stored secret. He reports 99% of functions present and 83% byte exact, using 14 Luna and 2 Opus 5.5 agents in the final weeks. Other lessons: lower the compaction threshold to 42%, and have a cron job re-inject the instruction document every hour. These are the author’s results; we did not check the project. He does not name the game.

Author’s post  ·  HN discussion

🔬 Can the agent do the research?

Epoch AI’s David Owen asked agents to invent a better post-training method than a GRPO baseline for Qwen3-8B, without being told about the human paper that already did it. The agents had a sandbox, GPU jobs on Modal and a budget of 3,000 GPU-hours across up to 50 GPUs.

Scores run from 0% at the GRPO baseline to 100% at the original paper’s performance. Epoch reports Claude Fable 5 at 2%, with a 90% interval of 0 to 10%. GPT-5.6 Sol scored 15% (2 to 28%). The original method re-run in the same setting scored 94% (72 to 100%). These are early results from a small number of runs. Epoch says successors Claude Fable 5.1 and GPT-6 Astra showed awareness of the task, so memorization affects newer tests. It also says grading is not fully automated and uses human review.

Epoch report  ·  HN discussion