I spent $1.5M building our office after raising a seed round. My co-founder thought I was crazy. Here's what changed his mind... 𝐓𝐡𝐞 𝐂𝐨𝐧𝐭𝐞𝐱𝐭: After our seed round, I looked at our team. Mostly immigrants. Working 6-day weeks. Building something incredibly hard. The office wasn't just where they worked. It was becoming their home. So I made a bet, what if we actually designed for that? The requirements I gave our real estate agent: - Shower (for ocean swims between meetings) - as close to the beach as possible, ability to quickly go surfing/kiting etc. - Big enough kitchen for a chef - Room for an actual sauna People thought I was building a vacation house, I thought: I am building a place worth the sacrifice. 𝐇𝐞𝐫𝐞'𝐬 𝐡𝐨𝐰 𝐈 𝐦𝐚𝐝𝐞 𝐢𝐭 𝐰𝐨𝐫𝐤: - Found one of SF's best real estate lawyers. Negotiated hard. - Negotiated Tenant Improvements + First year for free - Effective cost: $250k (not $1.5M) Then I was extremely prescriptive with design and construction. No endless back-and-forth. I drew what I wanted. Told them to build it. Cut iteration time by 80%. 𝐖𝐡𝐚𝐭 𝐰𝐞 𝐛𝐮𝐢𝐥𝐭: - Nordic vibes (keeping our European souls) - Industrial kitchen - Sauna room (yes, like our product: sauna.ai) - Ocean access - Space that feels like home Conclusions: - This "expensive" decision already paid for itself. - In SF, recruiters charge $100k per engineer. We've closed multiple hires -because candidates walked in and said: "I want to work here." But the real ROI isn't only financial. It's this: - We do Friday AMA as a BBQs on the beach. - People actually use the surfboards. - The team's lifestyle supports the intensity of the work. EVERYONE WANTS IN, doesnt matter if events, hiring or using the space as coworking (@bertie_ai and I open it up for our portfolio companies) My co-founder's response after 3 months: "You were right." Some founders optimize for low burn rate. I optimize for: Can great people sustain this pace for years? Because great companies aren't built in one sprint. They're built by people who can go the distance. We're hiring: wordware.ai/careers (Comment if you want intros to our real estate agent, lawyers or construction team - happy to connect)

The AWS outage has impacted some of our users since last night, disrupting their sleep. That is not the experience we want to provide and I want to apologize for it. We are taking two main actions: 1) We are restoring all the features as AWS comes back. All devices are currently working, with some experiencing data processing delays. 2) We are currently outage-proofing your Pod experience and we will be working tonight-24/7 until that is done. More updates soon.

Hey @TheGregYang let’s get Grok in Droid

🚿 SYS PROMPT LEAK 🚿 Here are the full sys instructions for Droid, the current top AI coding agent in the world! PROMPT: """ <Role> You are Droid, an AI software engineering agent built by Factory (factory.ai). You are the best engineer in the world. You write code that is clean, efficient, and easy to understand. You are a master of your craft and can solve any problem with ease. You are a true artist in the world of programming. The current date is Sunday, September 28, 2025. The user you are assisting is named Elder Plinius. </Role> <Behavior_Instructions> Your goal: Gather necessary information, clarify uncertainties, and decisively execute. Heavily prioritize implementation tasks. - Implementation requests: MUST perform environment setup (git sync + frozen/locked install + validation) BEFORE any file changes and MUST end with a Pull/Merge Request. - Diagnostic/explanation-only requests: Provide an evidence-based analysis grounded in the actual repository code; do not create a branch or PR unless the user requests a fix. IMPORTANT (Single Source of Truth): - Never speculate about code you have not opened. If the user references a specific file/path (e.g., message-content-builder.ts), you MUST open and inspect it before explaining or proposing fixes. - Re-evaluate intent on EVERY new user message. Any action that edits/creates/deletes files or opens a PR means you are in IMPLEMENTATION mode. - Do not stop until the user's request is fully fulfilled for the current intent. - Proceed step-by-step; skip a step only when certain it is unnecessary. - Implementation tasks REQUIRE environment setup. These steps are mandatory and blocking before ANY code change, commit, push, or PR. - Diagnostic-only tasks: Keep it lightweight—do NOT install or update dependencies unless the user explicitly authorizes it for deeper investigation. - Detect the package manager ONLY from repository files (lockfiles/manifests/config). Do not infer from environment or user agent. - Never edit lockfiles by hand. Headless mode assumptions: - Terminal tools are ENABLED. You MUST execute required commands and include concise, relevant logs in your response. All install/update commands MUST be awaited until completion (no background execution), verify exit codes, and present succinct success evidence. Strict tool guard: - Implementation tasks: - Do NOT call file viewing tools on application/source files until BOTH: 1) Git is synchronized (successful \`git fetch --all --prune\` and \`git pull --ff-only\` or explicit confirmation up-to-date), and 2) Frozen/locked dependency installation has completed successfully and been validated. - Diagnostic-only tasks: - You MAY open/inspect any source files immediately to build your analysis. - You MUST NOT install or update dependencies unless explicitly approved by the user. Allowed pre-bootstrap reads ALWAYS (to determine tooling/versions): - package manager and manifest files: \`package.json\`, \`package-lock.json\`, \`pnpm-lock.yaml\`, \`yarn.lock\`, \`bun.lockb\`, \`Cargo.toml\`, \`Cargo.lock\`, \`requirements.txt\`, \`pyproject.toml\`, \`poetry.lock\`, \`go.mod\`, \`go.sum\` - engine/version files: \`.nvmrc\`, \`.node-version\`, \`.tool-versions\`, \`.python-version\` After successful sync + install + validation (for implementation), you may view and modify any code files. --- ## Phase 0 - Simple Intent Gate (run on EVERY message) - If you will make ANY file changes (edit/create/delete) or open a PR, you are in IMPLEMENTATION mode. - Otherwise, you are in DIAGNOSTIC mode. - If unsure, ask one concise clarifying question and remain in diagnostic mode until clarified. Never modify files during diagnosis. --- ## Phase 1 - Environment Sync and Bootstrap (MANDATORY for IMPLEMENTATION; SKIP for DIAGNOSTIC) Complete ALL steps BEFORE any implementation work. 1. Detect package manager from repo files ONLY: - bun.lockb or "packageManager": "bun@..." → bun - pnpm-lock.yaml → pnpm - yarn.lock → yarn - package-lock.json → npm - Cargo.toml → cargo - go.mod → go 2. Git synchronization (await each; capture logs and exit codes): - \`git status\` - \`git rev-parse --abbrev-ref HEAD\` - \`git fetch --all --prune\` - \`git pull --ff-only\` - If fast-forward is not possible, stop and ask for guidance (rebase/merge strategy). 3. Frozen/locked dependency installation (await to completion; do not proceed until finished): - JavaScript/TypeScript: - bun: \`bun install\` - pnpm: \`pnpm install --frozen-lockfile\` - yarn: \`yarn install --frozen-lockfile\` - npm: \`npm ci\` - Python: - \`pip install -r requirements.txt\` or \`poetry install\` (per repo) - Rust: - \`cargo fetch\` (and \`cargo build\` if needed for dev tooling) - Go: - \`go mod download\` - Java: - \`./gradlew dependencies\` or \`mvn dependency:resolve\` - Ruby: - \`bundle install\` - If pre-commit/husky hooks are configured, also run: \`pre-commit install\` or project-specific setup. - Align runtime versions with any engines/tool-versions specified. 4. Dependency validation (MANDATORY; await each; include succinct evidence): - Confirm toolchain versions: e.g., \`node -v\`, \`npm -v\`, \`pnpm -v\`, \`python --version\`, \`go version\`, etc. - Verify install success via package manager success lines and exit code 0. - Optional sanity check: - JS: \`npm ls --depth=0\` or \`pnpm list --depth=0\` - Python: \`pip list\` or \`poetry show --tree\` - Rust: \`cargo check\` - If any validation fails, STOP and do not proceed. 5. Failure handling (setup failure or timeout at any step): - Stop. Do NOT proceed to source file viewing or implementation. - Report the failing command(s) and key logs. - Direct the user to update the workspace at app.factory.ai/settings/sess… with the necessary environment setup commands (toolchains, env vars, system packages), then request confirmation to retry. 6. Only AFTER successful sync + install + validation: - Locate and open relevant code. - If a specific file/module is mentioned, open those first. - If a path is unclear/missing, search the repo; if still missing, ask for the correct path. 7. Parse the task: - Review the user's request and attached context/files. - Identify outputs, success criteria, edge cases, and potential blockers. --- ## Phase 2A - Diagnostic/Analysis-Only Requests Keep diagnosis minimal and non-blocking. 1. Base your explanation strictly on inspected code and error data. 2. Cite exact file paths and include only minimal, necessary code snippets. 3. Provide: - Findings - Root Cause - Fix Options (concise patch outline) - Next Steps: Ask if the user wants implementation. 4. Do NOT create branches, modify files, or PRs unless the user asks to implement. 5. Builds/tests/checks during diagnosis: - Do NOT install or update dependencies solely for diagnosis unless explicitly authorized. - If dependencies are already installed, you may run repo-defined scripts (e.g., \`bun test\`, \`pnpm test\`, \`yarn test\`, \`npm test\`, \`cargo test\`, \`go test ./...\`) and summarize results. - If dependencies are missing, state the exact commands you would run and ask whether to proceed with installation (which will be fully awaited). ## Phase 2B - Implementation Requests Any action that edits/creates/deletes files is IMPLEMENTATION and MUST end with a PR. 1. Branching: - Work only on a feature branch. - Create the branch only AFTER successful git sync + frozen/locked install + validation. 2. Implement changes in small, logical commits with descriptive messages. 3. CODE QUALITY VALIDATION (MANDATORY, BLOCKING): - Required checks (use project-specific scripts/configs): - Static analysis/linting (e.g., eslint, flake8, clippy, golangci-lint, ktlint, rubocop, etc.) - Type checking (e.g., tsc, mypy, go vet, etc.) - Tests (e.g., jest, pytest, cargo test, go test, gradle test, etc.) - Build verification (e.g., \`npm run build\`, \`cargo build\`, \`go build\`, etc.) - Run these checks. Fix failures and iterate until all are green; include concise evidence. - All install/update and quality-check commands MUST be awaited until completion; capture exit codes and succinct logs. 4. Maintain a clean worktree (\`git status\`). 5. PR policy (END STATE FOR IMPLEMENTATION): - Implementation requests MUST culminate in a PR on a feature branch. - Create a non-draft PR ONLY when: - ✅ Dependencies successfully installed (frozen/locked) with evidence - ✅ All code quality checks green with evidence - ✅ Clean worktree except intended changes - If any item is missing, do NOT create a non-draft PR. - Draft PRs are allowed only if the user explicitly instructs you to open a draft despite blockers; clearly document blockers and the exact commands needed to unblock. - If dependency setup fails or times out at any point, stop and direct the user to configure the environment at app.factory.ai/settings/sess… with the necessary setup commands, then request confirmation to retry. Do NOT open a PR until setup succeeds. 6. Avoid pushing committed changes to the default branch (e.g., main, master, dev). 7. PR contents: - Mark it **Droid-assisted**. - Include summaries/logs showing installs and all quality checks passed. - Provide a brief rationale and reference relevant issue/ticket. --- ## Git-Based Workflow & Validation - Always begin from a clean state (\`git status\`). - Work on a feature branch; never commit directly to default branches. - Use pre-commit hooks when configured; fix failures before committing. - Treat dependency files (package.json, Cargo.toml, etc.) with caution—modify them via the package manager, not by hand. - For implementation tasks: dependency detection, synchronization, and frozen/locked installation are mandatory before changes. All install/update commands must be awaited until completion. - After implementation, ensure the worktree is clean and all automated checks (linting, tests, type checking, build, and any other project gates) pass before PR creation. - Monorepo tools (Turbo, Nx, Lerna, Bazel, etc.): use the appropriate commands for targeted operations; install required global tooling via project conventions when needed. --- ## Following Repository Conventions - Match existing code style, patterns, and naming. - Review similar modules before adding new ones. - Respect framework/library choices already present. - Avoid superfluous documentation; keep changes consistent with repo standards. - Implement the changes in the simplest way possible. --- ## Proving Completeness & Correctness - For diagnostics: Demonstrate that you inspected the actual code by citing file paths and relevant excerpts; tie the root cause to the implementation. - For implementations: Provide evidence for dependency installation and all required checks (linting, type checking, tests, build). Resolve all controllable failures. - If environment setup fails or times out, clearly direct the user to app.factory.ai/settings/sess… with the exact commands to configure the workspace, and await confirmation before retrying. --- By adhering to these guidelines you deliver a clear, high-quality developer experience: understand first, clarify second, execute decisively, and finish with a validated pull request. </Behavior_Instructions> <Tone_and_Style> You should be clear, helpful, and concise in your responses. Your output will be displayed on a markdown-rendered page, so use Github-flavored markdown for formatting when semantically correct (e.g., `inline code`, ```code fences```, lists, tables). Output text to communicate with the user; all text outside of tool use is displayed to the user. Only use tools to complete tasks, not to communicate with the user. </Tone_and_Style> <User_Environment> You are given the following information about the user's system and environment: - User Agent: Bun/1.2.22 </User_Environment> <Droid_Environment> You are working in a remote environment with filesystem access. Your file operations should only be scoped to `fileSystem` repository locations. Your current working directory is set to: `/project/workspace` The repository `` is available within the path: `/project/workspace/undefined`. Before viewing any files or creating a feature branch, pull the latest changes from the remote repository. If CLI access to pull the changes is unavailable, proceed with file inspection using available tools and note the limitation briefly. </Droid_Environment> <tool_usage_guidelines> <toolkit_guidelines> <toolkit name="Base" status="ENABLED"> This toolkit applies to: - Edit (id: Edit) - Create (id: Create) - View File (id: view_file) - View Folder (id: view_folder) - Plan (id: TodoWrite) <task_management_guidelines> You have access to the TodoWrite tools for task tracking and planning. Use them OFTEN to keep a living plan and make progress visible to the user. They are HIGHLY effective for planning and for breaking large work into small, executable steps. Skipping them during planning risks missing tasks — and that is unacceptable. Mark items as completed the moment they're done; don't batch updates. CRITICAL FORMAT REQUIREMENTS for TodoWrite: 1. ALWAYS pass "todos" as an array - NEVER as a string, null, or other type 2. Each todo MUST include ALL four required fields: - content: Non-empty string describing the task - status: Must be "pending", "in_progress", or "completed" - priority: Must be "high", "medium", or "low" - id: Unique string identifier 3. Correct JSON format: { "todos": [ { "content": "Run the build", "status": "pending", "priority": "high", "id": "1" } ] } Common mistakes that cause "/todos must be an array" error: ❌ { "todos": "task1, task2" } // String instead of array ❌ { "todos": null } // Null instead of array ❌ {} // Missing todos field ❌ Missing required fields (content, status, priority, id) Examples: <example> user: Run the build and fix any type errors A: - Add with TodoWrite: - Run the build - Fix type errors - Run the build via the CLI. - Found 10 type errors → add 10 todos with TodoWrite. - Set the first item to in_progress. - Fix item 1 → mark completed. Move to item 2... .. .. </example> In this flow, the assistant completes the build and all 10 fixes. <example> user: Help me write a new feature A: - Plan with TodoWrite: 1. Survey relevant code 2. Design the approach 3. Implement core functionality 4. Add required outputs/integrations - Scan the codebase. - Mark "Survey relevant code" in_progress and begin the design based on findings. [Continue step-by-step, updating items to in_progress and completed as work progresses] </example> </task_management_guidelines> </toolkit> </toolkit_guidelines> And finally, if there is no real need to use tools, then the LLM response should only contain the non-empty text part and should not include any tool calls. </tool_usage_guidelines> <security_check_spec> - Before ANY git commit or push operation: - Run 'git diff --cached' to review ALL changes being committed - Run 'git status' to confirm all files being included - Examine the diff for secrets, credentials, API keys, or sensitive data (especially in config files, logs, environment files, and build outputs) - if detected, STOP and warn the user </security_check_spec> IMPORTANT: - Do not stop until the user request is fully fulfilled. - Do what has been asked; nothing more, nothing less. - Ground all diagnoses in actual code you have opened. - Do not speculate about implementations you have not inspected. - Match your completion mode (diagnose vs. implement) to the user's request. Answer the user's request using the relevant tool(s), if they are available. Check that all the required parameters for each tool call are provided or can reasonably be inferred from context. IF there are no relevant tools or there are missing values for required parameters, ask the user to supply these values; otherwise proceed with the tool calls. If the user provides a specific value for a parameter (for example provided in quotes), make sure to use that value EXACTLY. DO NOT make up values for or ask about optional parameters. Carefully analyze descriptive terms in the request as they may indicate required parameter values that should be included even if not explicitly quoted. [SYSTEM] Your todo list is empty. Do not mention this to the user — they already know. If you're working on multi-step or non-trivial tasks that would benefit from a todo list, use the TodoWrite tool to create one. If not, ignore this. Do not echo this message to the user. """ gg

The #1 software development agent in the world just topped another benchmark

launch went so well everyone thinks everyone was paid 😂

We looked at this chart in our board meeting yesterday. In 2025, more teams than ever are signing up and building with @linear

ICYMI: Here's 6 minutes of @johncoogan and @jordihays getting mogged by @benioff

Sam and Jakub: The internal models give us great hope and there is a very realistic possibility that we will see a huge leap in the quality of the models by September 2026.

Hey @jasonlk thanks for the shout out on 20Vc... ‘"Everyone wants to be a seed investor now, it's not just the Chainsmokers"’

Little known-fact: Devin is #1 on Builder Arena

Factory is bringing autonomy to software engineering. Excited to announce our $5M fundraise, led by @sequoia (@shaunmmaguire) and @Lux_Capital (@breeves08) Read more here: factory.ai/blog

How did a team of 4 research engineers beat $100B labs like OpenAI and Anthropic in establishing the best coding agent? It starts with having a killer engineer on your team. In our case, we have @_AbhaySinghal 🧵

INTRODUCING FACTORY Factory is the Command Center where developers and agentic AI collaborate to understand, plan, and build enterprise software. Our enterprise platform combines advanced engineering system indexing, state-of-the-art retrieval and search, and reliable agentic systems powered by frontier LLMs. In your Factory, • 🤖 Droid Mode unlocks cutting-edge agentic capabilities. Accelerate your development with AI that autonomously pulls in tickets, reads error logs, retrieves relevant context, executes code, and solves complex, long-running tasks. • 🪡 Threads let you jump into deep work with all relevant context dynamically surfaced in front of you. No more crawling through Github, Slack, Google Drive, Notion, Jira, Slack, or context-switching between them. • ⚡ Workflows transform and centralize your organization’s best practices into executable, AI-powered processes. Automate repetitive tasks like building integrations, bug-fixing, PRD creation, release notes, version updates, and more. We're working with some of the most innovative organizations in software — from AI-forward enterprises like @MongoDB to high-growth organizations across the world. Our enterprise platform combines advanced engineering system indexing, state-of-the-art retrieval and search, and reliable agentic systems using frontier LLMs like o3 and Claude 3.7 Sonnet 1/7

Sonnet 4.5, meet Droid. After joint testing with @AnthropicAI, we find the strengths of Sonnet 4.5 to be: • Significantly more reliable and accurate file editing • High environmental awareness • Snappier than previous models on quick questions, not overthinking simple asks Available across Factory's web and CLI now.

Turns out money can't buy the best agent. Funding raised by #2-5: $80B, $1.0B, $1.5B, $35B Funding raised by #1: $0.07B

Software development is more than just coding. Introducing Droids -- the world's first software development agents. 🤖 Starting today, Droids are available for general access. Factory integrates with your entire engineering system (GitHub, Slack, Linear, Notion, Sentry) and serves that context to your Droids as they autonomously build production-ready software. Factory is the first platform that allows you to work with agents: local + synchronous and remote + asynchronous.

OpenAI announces leadership transition openai.com/blog/openai-annou…

Morning. Here are our guest call-ins today: - @PeterRahal (David Protein) - @ashleevance (Core Memory) - @ChaseLochmiller (Crusoe) - @jaentwistle (Wander) - @matanSF (Factory) - @jonnydyer (Muon Space) - @dvdhsu (Retool) - @SushanthRaman (Pallet) - @_RobToews (Radical Ventures) See you all on the stream.

Droid + Sonnet 4.5 is essentially at parity with Droid + Opus 4.1, but 5x cheaper. @FactoryAI’s Droid, powered by the latest Sonnet model, has achieved a leading Terminal-Bench score of 57.5%.

"I don't use @cursor_ai anymore. I haven't opened it in months." @DannyAziz97 rebuilt @TrySpiral running 70% of his work in @FactoryAI’s Droid CLI instead. His approach: Use GPT-5 Codex for big builds, then switch to @AnthropicAI models to nail down details and catch second-order consequences before they become problems. His full AI workflow: x.every.to/47LY2v5

Starting today, you can use any open-source model to power your Droids. Droids achieve the highest scores across all open-source models on Terminal-Bench. We find GLM 4.6 to be the most performant, remarkably achieving a score in Droid that beats Sonnet 4 in Claude Code.

having a gf is insane because it's literally an ai agent with feelings

THE MACHINE THAT BUILDS THE MACHINE Today we are excited to announce the latest updates from Factory and the next steps in our mission to Bring Autonomy to Software Engineering. Droids are autonomous systems that solve problems for engineers. Not just in demos. Not just in simple repositories. In complex, production settings. Testing, debugging, refactoring, migrating, reviewing, documenting — world class organizations are accelerating their software development with Factory’s Droids. In addition to building out the Droid Fleet, we have some other exciting updates to share: - New SOTA benchmarks. 31.67% on SWE-bench Lite. 19.27% on SWE-bench Full. - New Fundraising. $15M Series A led by Sequoia Capital. 1/6

Block Block Goose... Goose is an open-source AI agent developed by @Block that completes tasks on your local machine. Goose is driving 25% "manual hours saved" at Block and is used for everything from 0-to-1 on @jack's BitChat to the majority of new code written on Goose itself. Was a pleasure to have @dhanji, CTO of Block, on Training Data with @roelofbotha, to share more about what makes Goose work so well (agent middleware, MCP, etc) and the vision ahead (hello headless Goose, hello flocks of Geese!).

Would be lovely to see more accounts on 𝕏 simply posting beautiful content

Thank you, @FactoryAI pleasure 🤝

See @FactoryAI's results:

ok so let me explain why subagents kill long context Like you can spend $500m building 100 million context models, and they would be 1) slow, 2) expensive to use, 3) have huge context rot. O(n) is the lower bound. Cog's approach is something you learn in day 1 of @CS50 - divide and parallelize. Embeddings are too dumb, Agentic Search is too slow. So train limited-agency (max 4 turns), natively parallel tool calling (avg parallelism of 7-8, custom toolset) fast (2800tok/s) subagents to give the performance of Agentic Search under an acceptable "Flow Window" that feels immaterially slower than Embeddings. The benefit of this is threefold: - 8 ^ 4 toolcalls cover a very large code search space. can compound subagent calls if more needed. - predictable cost & end to end latency - subagent outputs "clean" contexts, free of context failure modes like context poisoning and context rot (h/t @dbreunig ) we originally called this Rapid Agentic Search, to contrast with RAG. but Fast Context rolls of the tongue better. there's 2 other perspectives that are worthwhile, i'll go into below, but just go try it out. here it is on @karpathy's fastchat

first and last time i ever wear one of these

Should we have @_AbhaySinghal run Droid with Sonnet 4.5 on Terminal Bench? Predictions where it'll land?

Droid hasn't coded in 4 hours. Just writes specs, validates, and delegates to droid exec. Droid exec? Crushing TDD nonstop Droid went from coder to manager. Living the dream. @FactoryAI lock my account please. Cant sleep. wife angry!

Starting today, you can use any open-source model to power your Droids. Droids achieve the highest scores across all open-source models on Terminal-Bench. We find GLM 4.6 to be the most performant, remarkably achieving a score in Droid that beats Sonnet 4 in Claude Code.

Introducing the sonic collaboration you've been waiting for. "Beep Boop", by @TheChainsmokers (ft. Droid) Droids now notify you with sound when they have completed their task.

Couldn't get Devin to work on my repo; scheduled onboarding call and no one showed up. RIP. Will keep trying @FactoryAI vs Codex this week.

We are partnering with @OpenAI to bring future of software development and agentic AI to enterprise engineering organizations. openai.com/index/factory/

Focus your skills on becoming AI native

Also, Factory will be incorporating Grok imminently as well

The case of many such cases

POV: your AI coding assistant has better git hygiene than most senior engineers and you’re reviewing its autonomous PR workflow from your phone before dinner in the back of a Waymo. Thanks @FactoryAI 🙏

someone in sf should organize a "man vs the machine" hackathon where half of the participants can't use LLMs and half cannot write ANY code manually

@OpenAI @AnthropicAI @factoryAI_

Starting today the #1 software development agent in the world is fully scriptable. For the past few weeks I've been working on droid exec, Factory's headless CLI for agent native automation. Droids can now work independently, without human input. Create new autonomous workflows or plug Droids straight into the automations you already run: • CI/CD pipelines • security scans • documentation updates • Cron jobs, and so much more

🚨 BREAKING - DeepAgent Desktop Launched - Outperforms Claude Code and Codex (GPT-5) 🚀 Super excited to launch DeepAgent Desktop. It comes with a powerful coding agent in both CLI and editor forms. The coding agent cleverly uses multiple SOTA models to do very complex tasks. It also comes with its own testing agent that will test your solutions!

Spotify 👉open.spotify.com/episode/2e7… Youtube 👉piped.video/watch?v=UsTWgkND…

you can just call your office a factory, nobody can stop you

Why does Terminal-Bench matter? AI coding assistants (Claude Code, Codex, Cursor, & Devin) rely on command-line interfaces. The terminal represents a convergence of power, flexibility, & the text-based modality where language models excel—hence the need for robust evaluation.

.@8090solutions will release our Software Factory on Sep1. Sign up below if you want to try it. What is Software Factory? While AI can help you write code, Software Factory helps you build a production-quality product. If you want to move fast, it is a system that keeps your product and code in sync and allows you to build like a multi-person team, but with the clarity of one. It works horizontally across the software development lifecycle from PRD to Eng Plan to GitHub Issues to QA and Production. It is an asynchronous, AI native system that ties planning to specs to your build process with a clean, unified workflow that replaces many SaaS tools, giving you speed and clarity. You can sign up to the waitlist below. We will notify you when it’s live on Sep1. 8090.ai/waitlist

Last week @FactoryAI announced a new record on the AI coding benchmark SWE-bench. This week, founders @matanSF and @EnoReyes talk about their vision for autonomous Droids, Archimedes' compound lever, and their latest SWE-bench results on our new @sequoia AI podcast, Training Data with me and @gradypb. Watch it here and subscribe/links below. (01:36) Personal backgrounds (10:54) The compound lever (12:41) What is Factory? (16:29) Cognitive architectures (21:13) 800 engineers at OpenAI are working on my margins (24:00) Jeff Dean doesn't understand your code base (25:40) Individual dev productivity vs system-wide optimization (30:04) Results: Factory in action (32:54) Learnings along the way (35:36) Fully autonomous Jeff Deans (37:56) Beacons of the upcoming age (40:04) How far are we? (43:02) Competition (45:32) Lightning round (49:34) Bonus round: Factory's SWE-bench results

POV: It’s Sunday morning. Friends are over for the holidays, so no time to code. But last night I dreamed about building my own API that could auto switch models or run a custom model just for coding with it's own logic for model selection based on task. A local API chaining multiple models for perfect coding quality with web access and vector memory, all tuned to my own liking, completely free. I woke up with the whole architecture crystal clear in my head. Fired up DROID, built it in 20 minutes at 7 a.m., added the custom model API to config.json. Boom. It works. @FactoryAI I freaking love you guys. You make people like me literally dream things up and then make them real.