20 posts tagged with "agentic engineering"

Over the past year, I have spent more time than I expected looking at AI coding benchmarks. Every few weeks, it seems, a new model arrives with another impressive result. SWE-bench scores go up, leaderboards change, and for a brief period one model appears to have established a meaningful lead—until the next release changes the picture again. As someone involved in building software products, I understand why we pay attention. If I am going to put an AI model somewhere inside a software development process, I want evidence that it can actually perform software engineering tasks.

There was a year when “cloud” was the whole pitch​

I started selling technology in 1999. I built what I believe was Canada’s first cloud computing company, at a time when saying the word “cloud” in a boardroom still required a diagram.

Codalio Blueprint is a free, MIT-licensed plugin that adds 10 planning and review skills to popular coding agents. The flagship skill, prd-builder, runs three lenses (Product & Scope, Architecture & Data, GTM) and synthesizes a single PRD that resolves contradictions before anyone writes code. Later skills review what gets built — auth exposure, performance cost, silent regressions, and which tests are worth writing. Outputs land as files in your repo — not ephemeral chat — so builders read a durable source of truth.

Last updated: September 11, 2026. Reviewed weekly.

If you're picking three open-source coding tools today: BMAD-METHOD to plan before you build (the only planning tool here that runs in a browser with no terminal), Goose to do the building (the only mainstream coding agent with a real desktop app), and the Claude Code GitHub Action to review what the agent wrote. All three are OSI-licensed and all three were pushed to within the last 48 hours.

This page is written for founders who can't code and are building anyway. Every star count below was read on September 11, 2026 — the first complete audit this page has had. The last two runs were cut short by GitHub's rate limit; there's now a way around it, explained at the end.

Last week's four ownership changes have all held: Goose is settled at aaif-goose, OpenCode at anomalyco, OpenHands at OpenHands, PR-Agent at The-PR-Agent. All four were pushed to again this week, so the transfers look like housekeeping rather than abandonment.

Idea to plan​

Before you write code you need something to build against. This is the category most vibe-coded projects skip, and skipping it is why they stall at 70%.

GitHub Spec Kit — 135,587 stars, MIT. Turns an idea into constitution, spec, plan and tasks across 30+ agents. The most rigorous option, and it opens with uv tool install and a Python 3.11 requirement.

OpenSpec — 67,989 stars, MIT. Proposals, specs and task checklists before coding, with a local dashboard.

BMAD-METHOD — 52,900 stars, licence unresolved (see below). Runs agile agent roles from idea to working software, with ChatGPT and Gemini web bundles.

Task Master — 28,063 stars, licence unresolved, last pushed April 28. Breaks a PRD into ordered, dependency-aware tasks.

Backlog.md — 6,704 stars, MIT. A markdown task board inside your git repo, with a local kanban UI.

codalio-blueprint — 5 stars, MIT. This one is ours. Turns a rough idea into a written PRD.

The one we'd install first: BMAD-METHOD. It's the only planning tool here with a genuine no-terminal on-ramp — the web bundles run as ChatGPT Custom GPTs and Gemini Gems, so you can do the entire planning phase before installing anything. Spec Kit is more rigorous and has two and a half times the stars, but its install command loses exactly the reader this page is for. The honest limitation on BMAD: the web bundles cover planning only. The moment you start implementing you're back in a CLI.

BMAD's licence is still unresolved, two weeks on. The README badge says MIT; GitHub's detector still returns no assertion, which normally means the LICENSE file has been edited. It's still our top pick and we're still not claiming the licence changed — but two weeks is long enough that this isn't a blip. If you're bundling BMAD into something you sell, open the LICENSE file and read it yourself. We've changed the licence column from "MIT" to "unresolved" to stop implying a grant we can't verify. Same for Task Master.

Task Master is the one we'd now hesitate over. Unresolved licence, and last pushed April 28 — four and a half months, unchanged from last week. Inside our six-month window, so it stays, but it comes off the page at the end of October if nothing lands.

On our own tool, plainly. codalio-blueprint runs three lenses (Product & Scope, Architecture, GTM) and synthesizes one PRD rather than stapling three documents together. We think that's well-built. It's also five weeks old, has 5 stars and 2 forks — unchanged from last week — no external contributors, and no test of whether the PRD is any good beyond examples we wrote ourselves. Last week we reported three new stars. This week, none. That's what a flat week looks like. Only Claude Code has real install instructions, whatever the README implies. Don't pick it over BMAD on our say-so.

Building​

Superpowers (285,156, MIT) installs a full agent methodology as composable skills across ~14 hosts. mattpocock/skills (259,497, MIT) applies senior-engineer review and TDD. OpenCode (206,685, MIT) runs a terminal coding agent against any model provider. Anthropic Skills (175,796, no root licence) holds the official reference skills and spec. OpenAI Codex CLI (123,351, Apache-2.0) and Gemini CLI (106,917, Apache-2.0) both run local coding agents. OpenHands (87,418, MIT) gives an agent a browser, terminal and editor. Cline (67,832, Apache-2.0) plans then edits with approval steps. Context7 (61,882, MIT) feeds agents version-correct library docs. Goose (54,129, Apache-2.0) runs an autonomous agent with a desktop app. Continue (35,869), vercel-labs/skills (31,395), Serena (29,183), Vibe Kanban (28,055) and Kilo Code (27,262) round it out.

The one we'd install first: Goose. The only mainstream open-source coding agent with a real desktop application — you see a window instead of a terminal — and Apache-2.0 with any-LLM support means no lock-in to one vendor's pricing. OpenCode has nearly four times the stars and is better if you're comfortable in a terminal, but it assumes you already are. Honest limitation: the desktop app hides the terminal, not the concepts. Extensions and MCP configuration still expect developer vocabulary, and you'll hit that wall on day two.

Goose has been pushed to repeatedly since moving to aaif-goose, the licence is unchanged, and it's up 230 stars on the week. One week isn't a guarantee, but it's the evidence we said we'd go and look for.

Vibe Kanban hasn't been pushed to since April 24 — same clock as Task Master.

Two things worth knowing before you install from this group. "Open source" often means the wrapper, not the engine: Codex CLI is Apache-2.0 and useless without a paid OpenAI plan, and Gemini CLI's free tier is a Google account benefit that can change without the repo changing. And the Anthropic Skills repo still has no root LICENSE file — confirmed again this week. Licensing is per-skill, and the document skills are source-available rather than open source.

Reviewing and QA​

This is where non-technical founders are most exposed. An AI wrote your code; something other than the same AI should look at it.

Trivy (37,870, Apache-2.0) scans dependencies, containers and IaC. Playwright MCP (37,011, Apache-2.0) lets an agent click through your live app. Gitleaks (29,234, MIT) detects committed secrets. Semgrep (16,590, LGPL-2.1) scans for security bugs. PR-Agent (12,950, MIT) reviews pull requests. Claude Code GitHub Action (8,844, MIT) reviews when you mention @claude. Trail of Bits Skills (7,041, CC-BY-SA-4.0) adds professional audit skills. cc-safety-net (1,535, MIT) blocks destructive commands.

The one we'd install first: the Claude Code GitHub Action. Typing "@claude review this" on a pull request is the lowest-literacy way to get a real second opinion on agent-written code, and it's MIT with no paid tier. Pair it with Gitleaks — an AI reviewer will happily discuss your architecture while ignoring the API key you committed in week one. Honest limitation: free to install, not free to run. Every review burns API credits and there's no built-in spend cap.

PR-Agent looks healthy after last week's move out of the Qodo org — pushed to this week, up about a hundred stars. The worry we raised hasn't materialised.

Trail of Bits Skills are the real thing, written by an actual security firm, but CC-BY-SA-4.0 is a content licence with a share-alike obligation. Read it before bundling commercially.

Shipping​

Supabase (109,056, Apache-2.0), Docusaurus (66,226, MIT), Coolify (61,677, Apache-2.0), GitHub MCP Server (32,869, MIT), semantic-release (24,034, MIT) and Changesets (12,384, MIT). All six were re-read on September 11 — these are the entries that carried stale August numbers for two weeks — and all six were pushed to within the last four days.

The one we'd install first: Supabase. The one piece of shipping infrastructure a non-technical founder can genuinely operate alone: clickable console, real free tier, Apache-2.0 so you can leave with your data. Coolify is better once you outgrow it, but the hardest step happens before Coolify appears — you have to rent a VPS and SSH into it. Honest limitation: some hosted-platform pieces aren't in the Apache-2.0 repo, so self-hosting isn't feature-equivalent.

One warning on the GitHub MCP Server: it needs a personal access token, and the easy broad-scope token hands an agent write access to every repository you own. Scope it down.

What didn't make the list​

Aider — still maintained, still excellent, wrong for this audience: its whole interaction model assumes you think in git commits and diffs. 48,897 stars, Apache-2.0, last pushed May 22 (our first current read on it). That date is approaching four months, which is worth watching on a tool this widely recommended.

Dokploy — open-core presented as open source, and the API confirms it: no licence assertion. Apache-2.0 applies only outside a /proprietary directory, and the proprietary licence forbids production use without a commercial agreement. Coolify is genuinely Apache-2.0 throughout and gets the slot.

Qodo-Cover — abandoned, with an explicit "no longer maintained" notice; the successor is paid. Automated test generation remains a real hole with no good open-source answer.

gpt-engineer, Devika, Claudia, snarktank/ai-dev-tasks, coderabbitai/ai-pr-reviewer — dead, stale, or 404. Named rather than silently omitted, because several still rank near the top of listicles on star count alone.

How to get exact star counts without hitting GitHub's rate limit​

Worth sharing, because it broke this page's audit twice. GitHub's unauthenticated REST API allows 60 requests an hour, which doesn't cover a 35-repo page, let alone four pages. But every repository page embeds its own exact figure in the HTML as "stargazerCount": <n> — the same number the API returns, not the rounded "48.9k" the page displays. Reading that costs no API quota. That's how every figure here got a current date for the first time.

Frequently asked questions​

How often is this updated? Every week. Entries we cannot verify are removed or flagged rather than quietly kept.

Is this the full page? Yes — this page is the canonical living guide, and it is updated here every week.

What changed this week​

Fixed the thing that kept breaking: every star count is now current. All 35 entries read September 11, including the nine that had carried August 28 numbers for two consecutive runs.

Corrected our own entry, in the unflattering direction. codalio-blueprint is still at 5 stars and 2 forks — flat on the week, not rising as we implied last week. A spot-check earlier in this run misread it as 1 star; 5 is correct. We'd rather print the correction than let either number stand.

Hardened two licence columns. BMAD-METHOD and Task Master now read "unresolved" rather than "MIT" — neither resolves to a standard SPDX licence, two weeks running. We're not claiming either changed; we're refusing to keep printing a grant we can't verify.

Maintenance clocks, now dated: Task Master last pushed April 28, Vibe Kanban April 24. Both stay this month, both come off at the end of October if nothing lands.

Last week's ownership changes all look healthy — Goose, OpenCode, OpenHands and PR-Agent each pushed to this week under their new owners, licences unchanged.

Notable movers: mattpocock/skills +10,881, superpowers +3,554, OpenCode +2,995, spec-kit +2,177, codex +1,923, OpenHands +1,270.

Nothing incomplete this run. For the first time since this page launched, there's no "we couldn't verify this" list.

Most AI tools will give you a demo. The question founders and operators actually ask is different: can I put a complete product in front of customers without hiring a CTO, a tech lead, senior developers, or cloud engineers first?

In April 2026 we ran that experiment. Four pipelines built the same meal-planning application. Codalio ran with no senior engineering team in the loop. Prompt-only Claude Code did not. We then scored every build against 111 things a customer would expect to work.

AI has already absorbed the junior coding work. Seniors set the plan, the architecture, and the task. AI writes the atomic pieces of code that used to be the first rung on every engineering career. So the hard question for new engineers isn't "can you code?" — it's "can you think like a senior and direct AI?"

That gap — between graduating and being trusted to lead — is the career problem of this decade. Codalio's Workshops & Training program exists to close it.

The software industry spent eighteen months solving the wrong half of the problem.

AI made code generation nearly free. Copilot, Cursor, and a wave of prompt-to-app tools let anyone turn a sentence into a running application in minutes. We called it "vibe coding," and it was genuinely exciting. But it solved the generation problem while quietly introducing a far more dangerous one: generating the wrong thing, confidently, at scale.

The fix isn't better code generation. It's the layer that comes before the code — the product specification. This is the shift from vibe coding to agentic engineering, and it's the difference between a demo that wows on Tuesday and a product that survives Wednesday.

For fifteen years the advice to non-technical founders was the same: build the smallest, ugliest thing you can, ship it, and pray you learn something before the money runs out. The "minimum viable product." A deliberately embarrassing version one.

That advice made sense when engineering was scarce and expensive. It doesn't anymore.

The constraint the MVP was invented to work around — you can't afford to build the real thing yet — has quietly disappeared. A single founder can now stand up software that's genuinely enterprise-grade from the first release: scalable, secure, reviewable, and shaped like an actual business instead of a demo. Not because they learned to code, but because they can now direct a full engineering team that happens to be made of agents.

Everyone upgraded the engine. Nobody checked the map.​

The conversation in AI building has quietly flipped. A year ago, the question was "which model?" Now it's "which harness?" — the orchestration layer that takes one instruction and fans it out across dozens or hundreds of parallel agents, each chipping at a slice of the work.

The proof point everyone repeats is real and genuinely impressive: a 750,000-line codebase ported from one language to another at 99.8% test pass, in eleven days. The takeaway people drew from it was "the harness is the moat." The same model scores differently depending on the wrapper around it, so the wrapper is where the leverage lives.

Half right. The harness is leverage. But leverage multiplies whatever you point it at — and most founders are pointing it at a guess.

The demo isn't the product​

A founder showed me a Lovable build last week. Working login, a dashboard with three charts, a settings page that actually saved. Built in an afternoon. They asked if this counted as "agentic engineering" — the phrase their advisor kept using.

It didn't. And not because the tool was wrong, or the output was bad. The demo was genuinely impressive. The problem was that nothing the agent produced had been asked for in a way it could defend. The login worked because someone clicked through a happy path. The dashboard rendered because the mock data fit. The settings page saved because nobody tried to save anything weird.

The second a real user shows up with a real edge case, the whole thing folds. Not because the agent is bad at code — but because the agent was improvising the whole time. There was no spec. There was a vibe.