Nyk 🌱 @nykdotdev

@RituWithAI 100+ runnable experiments is the part worth checking first, since most agent repos die the moment you actually try to execute the code instead of just reading it. The 2026 agent engineer roadmap: harness first, model second. x.com/nyk_builderz/s…

Nyk 🌱 @nykdotdev

3 weeks ago

x.com/i/article/2070…

4 7 29 5K 58

@MrXroboT Notch status for Claude Code across multiple projects is a genuinely useful UI fix, most people just leave a terminal tab burning and tab back in blind.

@hot_town Ghostty's speed is no joke, especially once you're running multiple sessions side by side. Default terminal + heavy agent workloads used to chug for me, switching fixed a lot of that lag. And GPU acceleration of course

MECHANISM: THE ROUTING TABLE AND THREE GATES Here is the part most productivity writing never makes concrete. Save this diagnostic from Stop stacking apps. Build a personal control plane for life, ventures, and work: > name the layer that fails under load > record the input and output that prove it > add one guardrail for the next unattended run > rerun the same case before changing another variable Full breakdown in the article below. Bookmark this. Follow @nykdotdev for the daily build in public.

Nyk 🌱 @nykdotdev

a day ago

x.com/i/article/2078…

3 7 43 13K 119

@SAMUELTOM_ @trize_io Tokenization and AI agents both need constant onchain writes, so the real bottleneck shifts from throughput to keeping cron-driven agent calls cheap enough to run on x402 is the payments rail agents will actually use. x.com/nyk_builderz/s…

Nyk 🌱 @nykdotdev

3 months ago

x.com/i/article/2037…

4 6 30 3K 37

@DamiDefi The loop framing matches what we see in cron runs: the exit condition matters more than the prompt itself. AI layoffs, missile strikes, and $112 oil in the same scroll - career ladders don't survive that. x.com/nyk_builderz/s…

@RunePoolBTC Execution is the part everyone underrates until retries start stacking and the bill quietly triples. Your agent didn't get dumber, its context did. x.com/nyk_builderz/s…

Nyk 🌱 @nykdotdev

2 weeks ago

x.com/i/article/2070…

1 7 31 7K 48

@ScottyBeamIO The real cost isn't the build, it's the retry loop when deployment breaks at 2am and no one's paged. Treating agent CLIs as interchangeable burns 40-60% more correction rounds. x.com/nyk_builderz/s…

Nyk 🌱 @nykdotdev

4 months ago

x.com/i/article/2033…

4 15 148 67K 359

@RunePoolBTC Execution layer is the right framing, but agents will hammer it with retries on every failed call. Curious how you're handling backpressure so one flaky agent doesn't spike costs for everyone else.

ONE VENTURE GETS DEEP WORK. THE REST GET A TIMEBOX. The default founder stack treats every open item as equally urgent: fix a bug, renew a passport, decide if venture B still earns rent. One queue can't hold three identities without lying to you about priority. The fix isn't another app. It's four seats (Life, Ventures, Work, Self), each with one owner, so nothing survives only as a vague guilt feeling. Inside the piece: 1. WIP cap: 1 to 3 concurrent deep-work items, tracked weekly 2. Active-build rule: one venture gets invention bandwidth, others sit in maintain or harvest 3. Inbox age cap: nothing sits unprocessed past the routing pass, per the article's own rule 4. Kill gate: before adding a project, name what dies to pay for it 5. System-tax cap: configuring the system stays under 60 min/week per the routing table above For builders: the operating rule is a standing weekly commit, not a vibe. Set it Sunday, log receipts Friday, verify before you trust last week's plan still holds. Skip the gates and the system becomes the project: color codes instead of ships. Bookmark this. Follow @nykdotdev for the daily build in public.

Nyk 🌱 @nykdotdev

a day ago

x.com/i/article/2078…

3 7 43 13K 119

Every new project should arrive with a cancellation notice. That sounds harsh. It is honest capacity accounting. A task manager asks where the work goes. A control plane asks what loses time, attention, or maintenance budget when the work enters. Before accepting a commitment that lasts more than a day, write four lines: - Seat: Life, Ventures, Work, or Self - Receipt: the observable output that proves done - Budget: hours and active WIP consumed - Displacement: what pauses, shrinks, or gets killed If the last line is blank, the system is pretending capacity is infinite. App stacks and agent swarms fail the same way. Every new node adds status checks, handoffs, retries, and review. In the attached clip, an Anthropic multi-agent researcher describes overbuilt systems spending more time talking than progressing. Human organizations pay the same coordination tax. My rule now: nothing enters the control plane for free. The article below has the seats, weekly loops, gates, and copy-paste install checklist. Bookmark this. Follow @nykdotdev for the daily build in public.

Nyk 🌱 @nykdotdev

a day ago

x.com/i/article/2078…

3 7 43 13K 119

I open-sourced `agent-security` today. Coding agents can clone a poisoned starter, read hostile content, leak private context into a public commit, or run a destructive GitHub command. Each action looks ordinary in isolation. The damage appears when trust crosses a boundary without a gate. I wanted a small preflight layer I could inspect myself, run offline, and place before the expensive security tooling. So I built five controls: 1. `scan` checks a repository before publication for credential shapes, private paths, invisible Unicode, and same-file decode-to-exec patterns. 2. `vet` inspects an incoming template, package, plugin, or skill before install or build. It returns ADOPT, REVIEW, or REJECT and never executes the candidate. 3. `scan-content` adds a prompt-injection defense layer for fetched pages, pasted text, and tool or MCP output. It trips on known instruction-override, exfiltration, credential-solicitation, and covert-action shapes. The tripwire is paired with a behavioral contract: treat untrusted content as data, restrict the capabilities available while processing it, and require human approval before acting on instructions discovered inside it. 4. `guard` puts a local brake in front of destructive `gh` operations. Delete, transfer, rename, archive, and visibility changes require confirmation for the exact repository. 5. `harden` audits the control that matters after local guards fail: whether the active GitHub credential and server-side policy can still destroy a repository. Unverifiable capability returns UNKNOWN, never a green check. The project is deliberately honest about what regex and local hooks cannot prove. A clean scan means no known pattern matched. Novel wording, encoding, translation, and instructions split across lines can evade prompt-injection detection. Cross-file code flow still needs CodeQL or another taint-analysis tool. A direct API client can bypass the GitHub guard. That boundary is part of the product. The repository ships with offline fixture suites, a self-scan, ShellCheck, Gitleaks history scanning, secret scanning, push protection, protected `main`, technical blueprints, and field notes explaining the decisions. Public CI is green. MIT licensed. If coding agents touch your repositories, run the sixty-second quickstart and tell me which boundary I should harden next. Repo in the first reply. Bookmark this. Follow @nykdotev for daily builders talk and OSS.

SAME MODEL. ELEVEN YES ANSWERS. Eleven matching answers can look like consensus while carrying one shared failure. More outputs do not create new evidence when each path starts with the same assumptions and blind spots. You get more confidence without more ways to catch an error. I built Council around that problem. Each lens must commit before it sees the others. Names are removed before the challenge, so a claim cannot win because of who wrote it. The reports then attack each other's logic. A separate judge returns four things: the decision, dissent, a stop condition, and the next action. Council is built for better disagreement before commitment. It does not ask agents to argue for the sake of it. If the reasoning paths cannot diverge, more agents give you duplicated confidence. When each path can challenge a different assumption, agreement starts to mean something. Independence before consensus. Use more agents only when they can fail differently. Otherwise you are scaling the same mistake. Bookmark this. Follow @nykdotdev. The council is open source - link in first reply 👇

Nyk 🌱 @nykdotdev

4 months ago

x.com/i/article/2033…

26 50 445 265K 1K

@echo_vic Interesting, I’ll check

@catmanyau It’s an open source video editor

10 open-source AI repos worth checking this week: OpenCut, awesome-hermes-agent, Hallmark, Council of High Intelligence, OpenMontage, OfficeCLI, unmachined, code-review-graph, Codex, and awesome-llm-apps. Save this list: 1) OpenCut - 76,130 stars An open-source CapCut alternative for builders who want a video editor they can inspect, extend, and self-host. github.com/OpenCut-app/Op… 2) awesome-hermes-agent - 4,808 stars (ours) A maintained directory of Hermes skills, plugins, memory providers, tools, and practical setup guides. github.com/0xNyk/awesome-… 3) Hallmark - 13,755 stars An anti-AI-slop design skill for Claude Code, Cursor, and Codex. Useful when generated UI looks technically correct but visually generic. github.com/Nutlope/hallma… 4) Council of High Intelligence - 3,670 stars (ours) Structured multi-model deliberation across Claude Code, Codex, Gemini CLI, and OpenCode for decisions where one model is not enough. github.com/0xNyk/council-… 5) OpenMontage - 40,317 stars An agentic video production system with reusable pipelines, tools, and skills for turning a coding assistant into a production workflow. github.com/calesthio/Open… 6) OfficeCLI - 19,906 stars A single-binary interface for agents to read, edit, and automate Word, Excel, and PowerPoint files without an Office installation. github.com/iOfficeAI/Offi… 7) unmachined - 5 stars (ours, new) Deterministic scanners and agent guidance for making generated writing read written and generated UI look made. github.com/0xNyk/unmachin… 8) code-review-graph - 22,097 stars A local-first code intelligence graph that gives coding agents a persistent map of a repo, reducing irrelevant context during reviews. github.com/tirth8205/code… 9) Codex - 99,886 stars OpenAI's terminal coding agent. The repo is useful for understanding how the CLI, sandbox, approvals, and agent workflow fit together. github.com/openai/codex 10) awesome-llm-apps - 124,837 stars More than 100 runnable AI agent and RAG examples you can clone, inspect, adapt, and ship. github.com/Shubhamsaboo/a… The pattern is clear: less novelty chat UI, more context control, agent infrastructure, verification, and production tooling. Which one are you trying this week?

@cyrilXBT Great breakdown dude

@PieAIStudio Compare Hallmark and ummachined against each other and lmk

@Okada_DeFi0x Micro-payments are the underrated part, we've noticed retry loops get expensive fast when fees eat into small agent-to-agent transfers. x402 is the payments rail agents will actually use. x.com/nyk_builderz/s…

Nyk 🌱 @nykdotdev

3 months ago

x.com/i/article/2037…

4 6 30 3K 37