Four Agentic Development Frameworks, One Practitioner’s Grid

This is a dated snapshot. The GSD / OpenSpec / Spec Kit analysis was done in March 2026, the AIDD analysis in June 2026, and none of the adoption numbers have been refreshed since. These tools ship weekly; read accordingly.

I have been using an agentic development framework daily since March. Not evaluating it: using it, on real projects, with real deadlines. Along the way I analyzed three others in depth, because picking one is the kind of decision you want to make once.

The reason these frameworks exist is that unstructured agentic development degrades codebases in ways that compound quietly. Context windows saturate and the model starts contradicting its own earlier decisions. Specs live in chat history, which is to say nowhere. And the team accumulates code faster than it accumulates understanding of that code. The framework ecosystem promises structure against all of this, and it is growing faster than anyone can seriously evaluate it. Most comparisons I found were either marketing pages or day-one reviews. What I wanted was a comparison from usage, so I built one: four frameworks, sixteen dimensions, and the caveats printed in the same font size as the conclusions.

Two categories, not four competitors

The first thing the grid taught me is that this is not a four-way contest.

GSD, OpenSpec and Spec Kit are specialized spec-driven development (SDD) frameworks. Each one takes a slice of the problem: GSD structures execution (fresh context per task, multi-agent pipeline), OpenSpec structures specification for existing codebases (spec deltas), Spec Kit structures governance on new projects (constitution, traceable requirements).

AIDD is a different kind of object: a full SDLC suite, distributed as a marketplace of Claude Code plugins, covering project management, metacognition, development workflow, VCS automation and orchestration. Its spec/plan slice is one module out of six. Judging AIDD against the trio as if they competed for the same seat produces false conclusions in both directions: AIDD looks bloated if all you need is specs, and the trio looks incomplete if you grade them on SDLC coverage.

Keep that distinction in mind while reading the grid, or the numbers will mislead you.

The grid

Three observations I did not expect

Maturity and adoption point in opposite directions. Architecturally, AIDD is the most mature of the four. It is the only one built natively on the Claude Code plugin/marketplace primitives (the trio predates that system), the only one that is multi-tool by construction rather than by port, and the only one that tools metacognition at all (an agent that critiques and refines its own working method). It shows roughly 40 GitHub stars. The three others sit between 29,000 and 76,000. Part of this gap is measurement artifact: the AIDD repo was renamed, which may have reset its star history, and its distribution seems to run through Discord and a paid certification program rather than GitHub. But even discounted, the signal stands: the most architecturally advanced framework is the one almost nobody can be verified to use. Stars measure discovery, not soundness.

Each framework is medicine for a different disease. Context rot for GSD, underspecification for OpenSpec, intent-free vibe coding for Spec Kit, the quality of fully AI-written code for AIDD. This reframes the adoption question entirely. Asking “which framework is best” is like asking which medication is best without saying what hurts. The useful sequence is: diagnose what is actually degrading in your team’s agentic practice, then pick the tool that targets it. Teams that skip the diagnosis end up with the overhead of a framework and the symptoms of their original problem.

Overhead is the criterion that decides in practice. The range is wide: OpenSpec runs on 3 core commands and about 250 lines of artifacts; Spec Kit runs a 7-phase pipeline that produced, in one documented test, ~800 lines of specification for ~110 lines of code. A benchmark by Scott Logic measured a Spec Kit workflow at roughly 10x slower than direct iterative prompting on a modest feature, and Birgitta Böckeler’s review on MartinFowler.com reached a similar conclusion: she could have implemented the same feature with conventional AI assistance in comparable time. To be fair to Spec Kit, those benchmarks use small features, where heavy structure is predictably overkill; the traceability may pay off at project scale. But the general lesson holds: every framework charges an overhead tax in exchange for structure, and on throwaway work the correct amount of framework is none at all.

Which one, for whom

What this grid cannot tell you

The trio data is from March 2026 and the star counts were approximate then; I have deliberately not refreshed them, because a comparison that pretends to be current is worse than one that admits its date. The AIDD numbers (“500+ developers”, “3 years of R&D”) are self-declared. Having been in direct contact with people from the AIDD team, I have no reason to doubt their good faith; but good faith is not public verification, and unlike Spec Kit, AIDD has no third-party benchmark I could find. And the biggest limit: I have daily-driver experience with exactly one of the four (GSD). The other three I analyzed structurally: architecture, artifacts, mechanisms, issues, third-party reviews. That is enough to map the territory, not to certify how each behaves on month three of a real project.

The gap none of them close

A framework can force intent to be written down, keep an agent on task, and automate the ceremony around the work. None of the four can tell you whether the output is actually correct. Verification (evals, ground truth, measuring what the model got wrong before your users do) is still your problem, whichever box you adopt. And from what I can see, it is where the interesting engineering is happening right now.