DigestAvi's weekly digest, 2026
The week of Oct 17 to 24.
Oct 17 to Oct 24, 2026. 34 bookmarks saved this week, sorted into 5 topics.
AI agents and the software factory
15 bookmarks
-
GitHub - humanlayer/humanlayer: The best way to get AI coding agents to solve hard problems in complex codebases.
A Bun-powered Turborepo that runs AI coding agents on user machines, with a unified desktop/web/daemon interface for tasks, sessions, plans, and reviews.
Why it matters. your software factory is exactly this problem: getting agents to solve hard problems in real codebases with human review in the loop, and HumanLayer is a reference architecture for the task/session/review layer.
-
An agent workforce to scale your work. | GREP AI
Self-improving agents that learn from every run to automate repetitive work, with MCP-based setup and more efficient token use, backed by Y Combinator.
Why it matters. agent harnesses that improve with each run are core to your factory thesis; GREP is a competing approach to the learn-from-runs loop you are building.
-
GitHub - Generated-Computers/Locals: Run many AI agents at once on your Mac, up to 3x faster.
Runs many AI agents at once on a Mac, claiming up to 3x speedups over standard mlx-lm by parallelizing tool calls and long conversations, compatible with OpenCode, Cline, Aider, and Claude Code.
Why it matters. running agent fleets locally on Apple hardware is your day-to-day setup, and parallel agent throughput is the bottleneck your factory has to solve.
-
Intent
Large-scale agent coordination where every task is isolated in its own workspace so hundreds of agents can run in parallel without collisions, with a living spec keeping work aligned.
Why it matters. parallel agents without collisions is the orchestration problem at the heart of your software factory, and the living-spec idea mirrors your spec-first working style.
-
RubyLLM 2.x: What’s new for your agents
RubyLLM 2.x adds tool approvals, a caller-controlled agent loop, durable agents that persist in a database and resume after restarts, plus batches, fallbacks, caching, cost tracking, and an MCP client.
Why it matters. you build in Ruby first and agents second; a Ruby-native agent library with durable, resumable agents is directly usable in your Rails-based factory tooling.
-
Obie Fernandez (@obie) on X
Obie Fernandez built decision-docket, an open-source Artifact-based docket where each decision is a card with the question, options with costs and gains, a recommendation, and what happens if you do nothing, so agents stop leaving overnight questions in scrollback.
Why it matters. overnight agents asking questions you wake up to is a factory workflow pain you know well; a decision docket is a concrete pattern for your review layer.
-
rahul (@rahulgs) on X
Ramp converged on one set of rails for its agentic systems: background agents, async agents, and live chat, covering an accounting harness, a general coworker, a coding agent, OCR agents, and a router with tasks, sandboxes, tool hooks, and queuing.
Why it matters. one shared harness across agent types is the platform decision your factory faces; Ramp's rails are a real-company reference for that consolidation.
-
Hiten Shah (@hnshah) on X
Hiten Shah on harnessification: AI harnesses that access tools, manage context, and retry are letting software work autonomously, shifting distribution power to whoever controls the harness, with Lovable, Replit, v0, and Salesforce Headless 360 as examples.
Why it matters. harness-as-distribution is the strategic frame for your factory and for Innovent's portfolio work; it names why the harness layer, not the model, may be where value accrues.
-
Avi Chawla (@_avichawla) on X
Jev plus Beacon, an open-source memory layer that captures agent sessions across harnesses into a shared index, dramatically cuts the cost of finding reusable corrections and trajectories across agents.
Why it matters. shared agent memory across harnesses is a missing piece in your factory; Beacon's cross-agent trace index is a pattern worth stealing for your own agent tooling.
-
Evaluating Agents on Highly Subjective Tasks
How to reliably evaluate agents on highly subjective tasks, with practical tools for getting robust results when judgments are noisy and subjective.
Why it matters. you care about measurable outcomes and evaluation before scaling agents; subjective-task evals are the hard part of judging factory output quality.
-
Hamel Husain (@HamelHusain) on X
Testing decision models for automated evaluations on 100 apartment-leasing traces, comparing F1 and cost, and showing how tuning the decision threshold beats the default 0.50 while holding precision and recall at 80% each.
Why it matters. threshold tuning for agent evals is exactly the kind of concrete, measurable technique your factory's verification layer needs.
-
Cutting our inference bill by 68% by being obsessive about optimization
Cloudflare cut AI code-review cost 68%, from about $2.46 to $0.77 per review, through prompt caching and smarter model routing, with median review time falling to 1 minute 32 seconds.
Why it matters. cost-per-agent-run is a factory metric you track; prompt caching and model routing delivering a 68% cut is a playbook you can apply to your own review agents.
-
Introducing Clairic, a new System One class of extraction models
Cloudflare's Clairic is a new System One extraction model that pulls, computes, estimates, and forecasts typed values from unstructured input, with higher accuracy, lower latency, and lower cost than traditional LLMs.
Why it matters. structured extraction from messy portfolio-company updates is the PostMoney problem; a cheaper extraction model class is directly relevant to that pipeline.
-
newsjack.sh — open-source PR skills for AI agents
An open-source PR toolkit for AI agents that spots newsjacking opportunities, finds angles, vets reporters, and pre-roasts pitches, with 31 skills as plain Markdown files.
Why it matters. agent skills as plain Markdown is your own convention, and newsjacking tooling fits your interest in growing products and personal brand without spammy output.
-
I built a software factory and gave a talk about it at XO Ruby Conf. One piece of feedback I got was that I didn't provide a demo. Well here is that demo folks were asking for. More to come soon… | Landon Gray
Landon Gray shares the demo of the software factory he built and presented at XO Ruby Conf, after attendees asked for one.
Why it matters. a fellow Rubyist's software factory demo is the closest peer reference to your own factory project; the demo format itself is useful for how you present yours.
Ruby and Rails
4 bookmarks
-
Ruby regexp security fixes: a newsflash
Ruby 4.0.9, 3.4.12, and 3.3.13 fix Onigmo regex bugs tied to four GitLab CVEs, including a double free and integer overflow that could enable remote code execution; JRuby and TruffleRuby are unaffected.
Why it matters. you ship Ruby in production, so a regex-level RCE in MRI is a patch-now item, and the Ruby Weekly framing is how you track these.
-
How Rails made its enums 4.5x faster
How Rails made its enums 4.5x faster, plus Aaron Patterson's static analysis for uninitialized Ruby variables, Ruby 4.0.8 notes, and updates to Net::SSH, RuboCop Spellcheck, and Passenger.
Why it matters. Rails performance internals are your home turf; a 4.5x enum speedup is the kind of core change worth understanding before it lands in your apps.
-
GitHub - seuros/psychowl: It knows what language you speak. Language detection for Ruby (pure Ruby + optional Rust engine) and Rust.
Psychowl is a language- and script-detection library for Ruby with a pure Ruby engine and an optional Rust engine sharing the same data and results, with parity tests between them.
Why it matters. a Ruby library with a pure-Ruby plus Rust-engine split is your kind of architecture question: when the native extension is worth it and how to prove parity.
-
GitHub - modosc/cloudflare-rails: fix request.ip and request.remote_ip in rails when using cloudflare
A gem that fixes request.ip and request.remote_ip in Rails behind Cloudflare by validating Cloudflare IPs and trusting X-Forwarded-For for the originating address.
Why it matters. you run Rails behind Cloudflare, so correct client IPs are a real, recurring production detail, and this gem solves it cleanly.
Security, secrets, and infrastructure
4 bookmarks
-
1Password Connect - 1Password Developer
1Password Connect servers let apps and cloud infrastructure securely access vault items via a private REST API, caching data locally so re-requests are unlimited after the first fetch.
Why it matters. secrets for agents and infrastructure is a live design question for your factory; Connect is the self-hosted pattern for giving apps vault access without scattering keys.
-
1Password for Agents: Secure AI Password Manager | 1Password
1Password for Agents lets personal AI agents log in to services without ever seeing your passwords, with integrations for Muse, Claude, and ChatGPT, plus vault storage for API keys.
Why it matters. agents that need logins without seeing passwords is the security model your factory needs before agents touch real accounts.
-
Secretive
Secretive is a native Mac app that protects SSH keys in the Secure Enclave, preventing export and avoiding on-disk key files; Secretive 3 adds macOS Tahoe support and post-quantum MLDSA keys.
Why it matters. you live on Apple hardware and manage SSH keys daily; Secure Enclave-backed keys with no on-disk material is the setup your own machines should match.
-
GitHub - cloudflare/security-audit-skill: A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
A coding-agent skill from Cloudflare that runs multi-phase security audits with independently verified, machine-readable findings, using ledgers and JSON schemas to track coverage.
Why it matters. agent-run security audits with verified, machine-readable findings fit your factory's review-layer design and your insistence on deterministic verification over LLM vibes.
Product, marketing, and growth
3 bookmarks
-
Nick Abraham (@NickAbraham12) on X
Lavender analyzed 28.3 million emails to find what drives replies; the post itself is mostly X login prompts rather than the study findings.
Why it matters. email reply-rate data matters for Clever Cat Prep outreach and for any product launch; the caveat is the findings themselves are not actually in the saved page.
-
Happenstance (@HappenstanceAI) on X
Happenstance is pitched as LinkedIn meets Claude meets ChatGPT, used to find precisely the right people inside a network of over 800,000 contacts.
Why it matters. investor-startup matching is VCmatch.ai's whole thesis, so a network-search tool that finds the right person is a competitive reference worth noting.
-
Alpha School is developing a book-rich, knowledge-rich curriculum. Ask me how I know.
Alpha School's Learning Science team, with compressed two-hour learning blocks, AI-personalized apps, and a closed-loop measurement system, is building a book-centered, knowledge-rich curriculum for K-2.
Why it matters. you founded Flatiron School and think deeply about technical education; Alpha's AI-plus-books model is a provocative counterpoint to pure app-based learning.
Tools, craft, and curiosity
8 bookmarks
-
Create a GitHub Issue - GitHub Marketplace
A GitHub Action that creates issues from template files using YAML front matter and Nunjucks templating for title, assignees, labels, milestones, and type.
Why it matters. templated issue creation is factory plumbing; automating well-formed issues from workflows is exactly the kind of small leverage your factory runs on.
-
GitHub - Gathin23/cleanyourtv: Debloat any Android or Google TV over Wi-Fi and get an Apple TV-style home.
Cleanyourtv debloats Android TV/Google TV over Wi-Fi with ADB and installs an Apple TV-style launcher, with no root and a one-line restore.
Why it matters. a self-hosted, no-root debloat with an easy restore is your kind of weekend project: practical, reversible, and assistant-driven.
-
Mechanical watch · Visual Learning
An interactive explanation of how a wound mainspring powers a mechanical watch, with a balance wheel and escapement regulating time at 28,800 beats per hour.
Why it matters. you collect visual, interactive explanations of how machines work; this one is a beautifully compact example of the genre.
-
Elvar (@elvar_official) on X
A free Claude Code skill called anatomy turns any explanation into an interactive machine visualization with step-by-step buttons, live measurements, and sound-word cues.
Why it matters. a Claude Code skill that generates interactive diagrams is directly in your agent-skills wheelhouse, and the format is one you could adapt for teaching.
-
Obie Fernandez (@obie) on X
Obie Fernandez made four AI-generated music videos in 18 days using Claude Code to script prompts, assemble footage, and display on-screen lyrics, a genre he calls slopcore.
Why it matters. you DJ and produce in Ableton and follow Obie's Ruby work; AI music-video production at this speed is a creative workflow worth watching, slopcore label and all.
-
The revised rules of engineering management with Will Larson
Will Larson on how the unwritten rules of engineering management have changed since the 2010s, from the ZIRP era to today, and what that means for the engineer-to-CTO path.
Why it matters. you have led product and engineering at CTO level; Larson's revised management rules are a useful check on how leadership expectations have shifted.
-
Dessa - the Conversation Model trained only on speech
Dessa is a conversation model trained only on speech; the saved page itself is Block's privacy notice describing data collection for the Dessa services.
Why it matters. speech-only conversation models are a novel AI direction, though the saved page is a privacy notice rather than the model details, so the substance is thinner than the title promises.
-
VoidZero (@voidzerodev) on X
tsrs is a Rust port of the TypeScript compiler, claiming up to 8x faster than TS 7.0 with about 70% less memory, running in WebAssembly and joining the Oxc toolchain.
Why it matters. compiler performance in Rust and WebAssembly is the kind of deep tooling win you appreciate, even outside your primary Ruby stack.