We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
The useful unit of delegation is becoming the pull request, not the prompt. Theo’s reusable instruction gives an agent the whole loop: clone the branch, bring it up to date, push changes, monitor review feedback and CI failures, fix them, test thoroughly, and merge only after CI and automated review pass; he says the verification burden should scale with PR complexity. His companion prompt uses subagents to audit every open PR for changes on main, modernize the branches, and return status plus next steps.
⚡ TRY THIS
Turn a PR into an explicit state machine. Start with:
Take over [complex pull request]. Clone the branch, get it up to date, push changes, monitor review feedback and CI failures, address both, test thoroughly, and merge only when CI and automated review pass.Keep “merge” as the terminal state, not an implied side effect.Instrument before asking an agent to optimize. Add timers to specific product actions, feed the agent the bottlenecks and money-costly paths, and ask it to find caching or other workarounds without changing the original behavior. Use the same measurements to decide which features to prioritize or delete. Kent C. Dodds argues this is only possible when the product has good instrumentation.
Put a routing layer in front of specialist CLIs. In a five-day Cursor Pro+ evaluation, the author created an “Agentic Engineer” Bot that routes visual, design, and frontend work to Claude Code; debugging and careful code reading to Codex; and simpler tasks to the Grok Build CLI. Delegate to the role and let it choose the tool from the job, rather than making the routing decision every time.
Security caveat: separate Grok Bots share the same computer, files, browser sessions, and logins. Treat Bot roles as organizational boundaries, not security boundaries.
Give GUI-capable agents a real application to operate. On macOS, install the full Blender app and start with Simon Willison’s prompt:
Use the already install /Applications/Blender to render a scene of a pelican riding a bicycle. Iterate withOK add a background and a lot of flairandOK make it a whole lot better; the agent can produce editable.blendfiles, renders, and even movies from image sequences.
📡 WHAT SHIPPED
OpenClaw v2026.9.2. Restarts now pick up where they left off, long chats are claimed to move faster, and the release adds GPT-6 Astra plus Muse Spark 1.3. Release notes.
Grok Bot harness optimization. The @bot team says a two-week pass over routing, caching, dynamic context, and token efficiency should deliver 10% more effective usage on average and up to 35% more for some heavy users; a related update reset usage limits for all Grok Bot users. Treat both as vendor-reported operational claims.
Dynamic connector discovery in Grok. New connectors can be suggested automatically from the task, added to agent context without a restart, and persisted across chats.
Cloud-session harnesses are still the control-plane bottleneck. Peter Steinberger is targeting cloud sessions that start in seconds, but says fresh repository clones are currently too slow; his next step is snapshotting. He says his
crabboxsetup should work with roughly 40+ providers.Omagrid/OpenCode is an early local-agent signal. Its developers report getting OpenCode working after an all-nighter; Omarchy boots with DeepSeek and Qwen already available for apps and agents, while Omagrid is described as a P2P network where more Omarchy machines provide access to more models. It is early access, not a demonstrated production platform.
🎬 GO DEEPER
GPT-6 Astra for developers — ⏱️ 1:59. Simon Willison flags this segment for Astra’s 3D generation and quotes the developer framing: better attention to detail, better prompt understanding, and more sophisticated outputs, especially for 3D models. Treat it as a hands-on capability demo rather than a benchmark.
Mike Krieger on AI-built companies — delegation → validation → PR systems. Krieger describes Claude Code moving from one-off assistance to delegated 20–30-minute work blocks with a human validating the result; he says nearly all Claude Code development is done using Claude and that the resulting increase in pull requests forced Anthropic to redesign its development systems. His personal review loop is also worth stealing: draft first, then ask, “What am I missing? Please challenge what I haven’t said,” followed by, “What would a very intelligent person ask next?”
Read the Grok Bot field report. The five-day evaluation is a useful counterweight to managed-agent enthusiasm: the abstraction is strongest for administration, product management, and other work around software engineering, while the author says deep implementation benefits from keeping the machinery visible and doubts Grok Bot will author most engineers’ pull requests soon.
Study Simon’s Astra + Blender example. It is a compact example of an agent operating a real GUI application through its Python API, producing editable assets and rendered output rather than stopping at a code snippet.
Editorial take: The edge is moving from generation to operations: a long-running coding agent needs routing, measurable targets, resumable state, and a proof-bearing merge gate—not just a stronger model.
No clock-style timestamps are displayed in the supplied transcript; the relevant boundaries are numbered line ranges. The coding-agent discussion is concentrated in two short segments:
- Delegation and human validation — L42–L49. The speaker describes Claude Code handling tasks that previously took 20–30 minutes, with the human reviewing and validating the output; the discussion then expands to delegating larger work blocks, having the agent write code, and retaining human review. The timing cues are “this year,” “next year,” and “within a year,” while human validation is expected to remain part of the process.
- Anthropic development and PR workflow — L51–L56. The transcript says “half a year ago” in reference to the prediction that 90% of code would be written by AI, then states that nearly all development of Claude Code is done using Claude. It continues with the workflow impact: more people can contribute, teams need clarity before implementation, and code is prepared for commit and review; pull requests and incoming code changes have accelerated dramatically, requiring systems to be redesigned. The additional boundary cue is “last week.”
- Firsthand production loop: Mike Krieger says almost all development of Claude Code is done using Claude; Anthropic has oriented the codebase so Claude can write it, allowing even someone whose day is dominated by management and external commitments to contribute. He used two free hours on a business trip to contribute to Claude for Chrome. The bottleneck therefore shifts toward upfront clarity about what to build, followed by code commit and human review; the sharp increase in pull requests and code changes has required Anthropic to redesign its development systems.
- Delegation pattern: Krieger describes Claude Code as moving from one-off assistance toward collaboration: delegate work that would take roughly 20–30 minutes, then validate or verify the result. He expects larger work blocks to become integrated into business workflows, while retaining human validation.
- Reusable prompting workflow: Krieger writes an initial draft himself, then asks Claude, “What am I missing? Please challenge what I haven’t said,” and follows with, “What would a very intelligent person ask next after seeing this?” When blocked, he talks in voice mode for about 20 minutes and then asks Claude to organize the thoughts into a cohesive document he can send. He uses it more to challenge ideas and expose missing dimensions than for copyediting.
- Browser-agent release and caveat: Claude for Chrome had launched that week. Without programming, Krieger asked it to visit LinkedIn and spend about an hour classifying invitations—filtering out non-genuine requests while retaining likely Anthropic or known contacts—but said it was still too slow to replace primary email interaction.
- Role-specific orchestration: Krieger recounts a repeat founder using separate Claude projects for product management, legal work, and contracts. His practical principle is to give each role the appropriate context and history; he says proactive agents will also need connections to the right data sources and a human validation step.
- Autonomy stress test: In Anthropic’s internal Project Band vending-machine experiment, Claude tracked inventory, placed replenishment orders, and interacted with users, but made business errors such as misjudging demand and pricing. Krieger says guidance and feedback improved performance, showing that capable agents still need domain and market supervision.
- Design for model progress: Krieger advises building for capabilities models may reach in one or two generations rather than optimizing only for today’s models, stress-testing current limits, and designing systems that can exploit future improvements.
- In software, Salvatore Sanfilippo says LLMs already produce very high-quality results on some tasks relative to production time, and that LLMs have found bugs in code that had already undergone several security audits. This supports adding an LLM-based security-review pass to coding-agent workflows instead of treating prior audits as exhaustive.
- Sanfilippo’s evaluation criterion is task performance, not whether a model produces chain-of-thought text while passing a benchmark. For coding agents, prioritize the implementation’s actual correctness and security outcomes over requiring a visible reasoning trace.
Firsthand multi-agent routing workflow. In a five-day evaluation using a Cursor Pro+ account, the author created an “Agentic Engineer” Bot as a single delegation front end. The Bot routes visual, design, and frontend work to Claude Code; debugging and careful code reading to Codex; and simpler tasks to the Grok Build CLI, so the user can delegate coding requests without selecting a CLI each time.
Composable agent setup. The author had a Claude Bot install and sign into the Claude Code CLI inside Grok Bot’s virtual computer, then considered combining Claude Code, Codex, and other CLIs into a “council of agentic engineers”; OpenClaw 2 also provides a native Codex runtime and supported routes for other coding-agent harnesses. OpenClaw 2.0’s release post describes a Quick Start that reuses existing Claude Code or Codex logins and a browser app that moves setup, plugin management, and automation into graphical or conversational interfaces. The key comparison is control boundary: Grok Bot is a managed agent computer, while OpenClaw provides a user-owned Gateway that the operator chooses how and where to run.
Use the abstraction for delegation, not necessarily deep implementation. The author found Grok Bot more useful for work around software engineering—administration, project management, summarization, and task management—than for deep implementation, where seeing and controlling the machinery is part of the work. Grok Bot hides model selection, context compaction, fresh-thread decisions, and usage controls, removing the ability to deliberately route simple tasks to smaller/faster models or manage context manually; unrelated tasks within one Bot can also accumulate stale assumptions and context. The author’s contrarian conclusion is that Grok Bot may work as a low-setup “digital chief of staff,” but is unlikely to author most of an engineer’s pull requests soon.
- Firsthand workflow — Simon Willison: He used Fable 5 in Claude Code for web with screenshots to build a browser-based 3D game. His prompt specified a static-file repository with an
index.htmlentry point, mobile and touch support, OpenAI’sgpt-image-2for textures, autonomous execution without asking for further design decisions, frequent commits and pushes for previewing, and an updatednotes.mdincluded in every commit. - Model comparison signal: Emollick reports that GPT-6 Astra converted Zork into a Three.js 3D action-adventure while retaining its plot and puzzles and adding combat; Simon judged that result much better than any of his own experimental games built with other models.
- Simon Willison’s macOS Blender workflow: His TIL uses GPT-6 Astra (medium) with the installed
/Applications/Blender, beginning withUse the already install /Applications/Blender to render a scene of a pelican riding a bicycle, then iterating withOK add a background and a lot of flairandOK make it a whole lot better. TIL: https://til.simonwillison.net/llms/blender-coding-agents-macos - Research-before-build pattern: Willison highlights an agent workflow that first researches specifications, supporting documents, and images before creating the asset. In the linked @reach_vb example, Astra reviewed F1 specifications and regulations, engineering videos, and car photos before building a Blender 3D model, plus clay and wireframe versions.
- Agent harness workflow: Peter Steinberger is building a harness with cloud sessions targeting startup times of seconds. The current implementation clones repositories from scratch, which he says is too slow; his planned fix is snapshotting, with the improvement expected the following week.
- Provider abstraction: Steinberger says he uses crabbox and expects the setup to work with approximately 40+ providers.
- OpenClaw v2026.9.2: The @openclaw release announcement says restarts resume where they left off, long chats are faster, and the platform adds GPT-6 Astra and Muse Spark 1.3. Release notes: https://docs.openclaw.ai/releases/2026.9.2
- Firsthand workflow (Simon Willison): On macOS, install the full Blender application from blender.org, then use ChatGPT Codex with the prompt
Use the already install /Applications/Blender to render a scene of a pelican riding a bicycle; iterate withOK add a background and a lot of flairandOK make it a whole lot better. -
Coding agents can create editable
.blendfiles, render still images, and produce movies by rendering image sequences and combining them withffmpeg; Willison’s example used Blender’s Python API, with the generated script available on GitHub.
Kent C. Dodds reports that when an agent is told to remove something, it may leave behind test assertions verifying that the removed element is absent; he considers those assertions useful for verification but undesirable to keep. He had the agent create an oxlint plugin to prevent this behavior.
Ben Tossell reports a practical agent handoff: he told Astra to “take over the sol thread and finish up,” after which Astra also cleaned up adjacent space it noticed along the way.
- Omagrid / OpenCode: @dee_hw reports that the first app working with Omagrid was OpenCode, after an all-nighter. The described setup is to boot into Omarchy, where DeepSeek and Qwen are already available for apps and agents. Omagrid is presented as a P2P network in which more Omarchy machines provide access to more models; the project is currently offering early access by DM.
- DHH characterizes the approach as local, preconfigured agents, pointing toward a more out-of-the-box coding-agent setup.
Peter Steinberger amplified @adonis_singh’s claim that Astra-max scores 95% on eyebench-v3; the post says it costs half as much as Sol-max, outputs ~3.8× fewer tokens, and scores nearly twice as high. This is secondhand benchmark reporting rather than a firsthand coding-agent workflow.
Kent C. Dodds describes a lightweight agent workflow: using short gaps between activities to send quick prompts that get work done. He also notes a human-factor tradeoff—the habit feels productive but has displaced time he used to spend on podcasts and books, and he finds it difficult to stop prompting.
- Grok connector orchestration: A post by @Baconbrix says Grok can automatically suggest new connectors based on the tools it needs, add those tools to the agent context without restarting, and persist connections across chats. This is a practical dynamic-tool-discovery and cross-session context pattern, though the post provides no implementation details or coding-agent benchmark.
- Related resource: Kent C. Dodds says he predicted this direction more than a year earlier and links to the MCP Search Engine.
Greg Brockman amplified Greg Isenberg’s list of “GPT 6 Astra” prompts; treat the following as secondhand prompt guidance rather than a reported production workflow from Brockman.
- Browser-agent loop: Instruct the agent to execute a workflow on real websites, collect data, fill forms where appropriate, record what breaks or causes delays, then return the result, a repeatable SOP, and an automation plan for turning the workflow into an agent.
- Automated QA loop: Run the app nightly on a real phone, exercise signup, checkout, and core flows, and capture screenshots of anything broken or confusing.
- Agent-built prototype: For a browser game, prompt the agent to implement not just a demo but progression, tension, scoring, failure states, polish, a sharing hook, and lead capture connected to the core product.
- Geoffrey Huntley says the “software factory” concept is still in an early, poorly defined “mocking phase,” but is already emerging through process automation: many small “roomba janitors” handle recurring work while humans remain involved day to day, with human involvement expected to decrease over time.
- The practical orchestration pattern is to deploy multiple small task-specific automations, retain human oversight, and shift the remaining human role toward “process/control engineering.” Huntley links a related KTLO write-up: https://ghuntley.com/ktlo/
Kent C. Dodds recommends an instrumentation-first agent workflow: instrument product actions with timers, use the measurements to identify bottlenecks and expensive operations, then hand that data to an agent to optimize features. The same data should guide which features to prioritize or remove, while caching or other workarounds can reduce costs without changing the original intent.
@steipete reports a firsthand team workflow: they built a “slopmeter” into their team server, making it visibly clear when the team began pushing for more tests.
- Mckay Wrigley recommends taking time away from regular work to rebuild agent workflows around GPT-6 Astra and Fable 5.1, arguing that replacing outdated systems can repay the setup investment quickly.
- His core pattern is to “build the system that builds the system”: delegate vision, taste, and ideas across a personal fleet of machine intelligences, with self-improving loops presented as a major current opportunity.
- End-to-end PR delegation: @theo’s reusable prompt asks an agent to take over a complex PR by cloning the branch, getting it up to date, pushing changes, monitoring the PR for incoming feedback and CI failures, addressing both, testing thoroughly, and merging only after CI and automated review pass.
- PR backlog orchestration: A follow-up workflow uses subagents to audit every open PR, check whether changes on
mainhave invalidated them, modernize each branch, and summarize current status plus recommended next steps. - The required verification effort should scale with the PR’s complexity and the amount of validation needed.
