🛠️ DeepSeek's Agent Runs While You Sleep

DeepSeek shipped a terminal-native coding agent that runs on prefix-cache stability. In practice: you leave it running. It doesn't drift, stall, or balloon context. You come back and the work is done.

That is not a small claim. Most agents I've tested need babysitting. A long task means context pressure, which means drift, which means you're debugging the agent instead of the code. DeepSeek-Reasonix is engineered specifically around that failure mode. The prefix-cache architecture keeps state stable across long sessions. The community has noticed: 29,822 stars on GitHub since launch.

The real question isn't "is this good?" It's "what does a reliable overnight coding agent change about how you scope work?" If you can trust it to run a refactor while you sleep, your unit of productive work just got a lot bigger.

That's the lead story. Below: six repos total, three Claude Code skills worth stealing, one stack pick, and today's signals.

New here? This part's for you. (Operators, skip to The Drops.)

New here? This block is for you.

### What is fine-tuning?

You know how a new employee starts on day one knowing general office skills but nothing specific about your company? They get trained. Fine-tuning is that training process for an AI model.

A base model already knows a lot: how language works, how to follow instructions, general knowledge from the internet. Fine-tuning takes that model and runs it through a second round of training on a narrower dataset, your data, your style, your domain. After fine-tuning, the model behaves like someone who has been doing your job for a year.

The tradeoff: fine-tuning takes real effort, real data, and real compute. For most operators just starting out, prompting well gets you 90% of the way there. Fine-tuning is the next step when prompting hits its ceiling.

### Do this today

Open whatever AI tool you use daily. Write a prompt that includes a short example of the output you want, with your own words and style, before asking it to do the task. Compare that output against a prompt with no example. That gap you just saw? That is prompting doing some of the work fine-tuning does.

### How to use today's picks

Today's Drops include 6 repos and 3 Claude Code skills: a terminal coding agent with 29k stars, three Claude Code agent skills you can install directly, a vector database, and a real-time voice agent framework. For each one: read the one-line pitch below, click through, and skim the README (the project's front-page notes on what it does and how to start). Don't install anything yet. You're just building a mental map of what exists. Install when one of them solves a problem you're actually stuck on.

### Plain English

Fork, a personal copy of someone else's repo. You take their code, copy it to your own account, and can change it however you want without touching the original.

Dashboard, a visual screen (usually in a browser) that shows you live numbers, status, or logs about a running system. "Local-first dashboard" means it runs on your machine, not a remote server.

We read your replies. You told us you're here to learn, so we built this block for you. Hit reply with a word you want decoded and we'll add it to the queue.

[Affiliate] Runpod, Local inference is the whole point until the model outgrows the card in your machine. Runpod rents GPUs by the second, so you can size up for a long agent run and stop paying when it finishes. Worth a look if your own GPU is the ceiling. We may earn a commission.

[Repo] esengine/DeepSeek-Reasonix, a DeepSeek-native coding agent built for the terminal, engineered around prefix-cache stability. 29,822 stars. The design bet: a long-running agent that doesn't drift. Leave it on a refactor task, go do something else.

[Skill] addyosmani/web-quality-skills, agent skills for Lighthouse and Core Web Vitals optimization, authored by Addy Osmani. 2,577 stars. Wire this into your Claude Code session and get structured web performance audits without leaving the terminal.

[Skill] htdt/godogen, autonomous game dev for Godot, Bevy, and Babylon.js, using Claude Code and Codex under the hood. 5,218 stars. Describe the game mechanic; the agent generates the scaffolding. Genuinely useful if you are prototyping something interactive.

[Skill] lewislulu/html-ppt-skill, AgentSkill that builds HTML-based slide decks: 24 themes, 31 layouts, 20+ animations. 7,594 stars. No Canva, no Google Slides. Tell Claude Code what you want and it outputs a deployable presentation.

[Repo] modelcontextprotocol/modelcontextprotocol, the official MCP specification and documentation. 8,844 stars. If you are wiring MCP servers into any agent stack this week, this is the authoritative reference, not a third-party explainer.

[Repo] mikehasa/agentacct, shows you what your coding agents did and what it cost. Breaks each task into work steps: tools used, files changed, tests run, time and tokens spent. 545 stars. Local-first dashboard for Claude Code. The gotcha you will hit without this: a long agent session eats tokens you didn't budget for and you find out at the end of the month.

[Repo] microsoft/BitNet, official inference framework for 1-bit LLMs. 39,810 stars. If you are running models locally and want to cut memory and compute without sacrificing capability, this is the framework to test.

[Repo] milvus-io/milvus, high-performance, cloud-native vector database built for scalable ANN search. 45,490 stars. The go-to when your retrieval layer is the bottleneck and pgvector has hit its ceiling.

[Repo] livekit/agents, framework for building real-time voice AI agents. 11,906 stars. If your agent stack needs to handle a live voice interface, this is the most production-ready open option I have seen.

[MCP] FlanChanXwO/pixiv-cli, Pixiv in your terminal: a CLI, MCP server, and Go SDK covering discovery, accounts, creators, collections, and downloads. This passes the bar because it is a complete interface to a large creative asset platform, not a thin wrapper. The non-obvious use: wire it into a Claude Code session as an MCP server to pull reference imagery programmatically for a design or generative workflow. You get structured search and download without leaving the agent context.

- AWS embeds Superblocks into private clouds. AWS now lets enterprise customers run the vibe-coding tool Superblocks inside their own private cloud infrastructure. The operator consequence: the model-agnostic internal-tool layer is becoming a default AWS procurement category, not a side experiment. If you sell tooling into enterprise accounts, the incumbent just got a very large distribution partner. (TechCrunch)

- EU AI transparency rules go live today. The EU AI Act's labeling and transparency requirements are now in effect, including mandated disclosure for AI-generated content and deepfakes. If any of your products touch EU users or generate synthetic media, this is not a future compliance item. It is live. (The Verge)

- Vercel WAF for Blob is generally available. Vercel's Web Application Firewall for Blob storage is now production-ready on all plans. Existing rules carry over automatically. If you are storing agent outputs, user-uploaded files, or generated assets in Vercel Blob, you now have a WAF in front of it at no extra config cost. (Vercel)

- David Crawshaw is running Claude Code as a nightly cron. Simon Willison flagged Crawshaw's prompt: a nightly cron that fetches upstream changes, rebases local patches, verifies the software works, and auto-replaces the binary. That is a working autonomous maintenance loop. The pattern is real and stealable today. (Simon Willison)

The move: automated nightly rebase loop.

David Crawshaw's pattern from today's Signals is worth stealing directly. Here's the shape:

1. Write a single Claude Code prompt that does four things in sequence: fetch upstream changes, rebase your local patches on top, run your test suite, replace the binary if tests pass. 2. Schedule it as a nightly cron (crontab -e, pick a time when your machine is on or use a remote runner). 3. The prompt is the config. Keep it in a /prompts/nightly-rebase.md file in your repo so it versions with the code. 4. Add one guard: if the rebase hits a conflict, the agent should halt and write a CONFLICT.md note rather than forcing a bad merge. You'll wake up to a clear failure state, not a broken binary.

You'll know it worked when you open your terminal in the morning and the binary is newer than last night without you touching it.

The payoff: one less maintenance task that requires you to be awake and paying attention.

I was one decision away from reverting a change that was working, on the strength of a number that could not have told me either way.

Before you credit or blame a change, check whether the number was already moving for a reason you had nothing to do with.

Our open rates have been sliding, and the obvious suspect was the subject lines. We changed the approach, wrote to the new direction, sent, and the rate came in lower than ever. Read alone, that is a verdict: the new style is worse, revert it.

Then I lined up the last several sends against how many people each went to. The rate tracked recipient count almost perfectly, downward, send after send, regardless of what the subject line said. The list has grown substantially in a short window, largely through paid acquisition, and a newer subscriber opens at a lower rate than someone who has been reading for months. The rate was falling because the denominator was changing. The copy was along for the ride.

The honest test was to find a metric that dilution does not touch the same way. Clicks from confirmed human readers: the new direction ran meaningfully ahead of the old one. So the change is right, and I was about to reverse it on the strength of a number that could not have told me either way.

Two things I am keeping:

• A rate is a fraction, and a fraction can move entirely because of its bottom half. Always ask what happened to the denominator before you interpret the ratio.
• Match the metric to the mechanism. A subject line's job is to earn the click. Judge it on clicks, and stop promising yourself it will fix a number that is really about who is on the list.

What experiment did you call a failure before checking what else moved? Hit reply. We read every one.

Six repos, a nightly automation pattern you can ship before bed, and the EU compliance clock just started. The one thing worth doing tonight: pull agentacct and see what your agents actually cost. Most operators are surprised.

Before you go: we started a room for people actually building with agents. Post what you shipped, get unstuck, and drop what you want us to build next. Join the community →

See you tomorrow.