The toolbox in July 2026
A map of the stack, not a recommendation. Every category below had a clear leader in July 2026 and none of them are safe to assume in 2027. The durable finding across the whole page: the framework is rarely the differentiator; harness and context design isreputable.
One tool on this page is ours rather than the market’s. The dimmer is chapter 8’s four-property check (cheap · immediate · unfakeable · bounded) as a standalone instrument that holds its setting in the URL, for deciding one loop at a time whether the check it has today earns it the dark.
Spec-driven frameworks
The category that turns “vibe coding” into something an agent can be held to.
| Tool | Shape | Use when |
|---|---|---|
| GitHub Spec Kitprimary | Constitution of immutable principles, spec, plan, CLI; 30+ agents supported | Greenfield; you want the reference implementation |
| BMAD-METHODprimary | 12–21 specialised agents simulating an agile team, versioned artifacts throughout | You need the whole team simulated, and can absorb $800–2,000/dev/monthreputable |
| OpenSpecprimary | Markdown, tool-agnostic; every change is a proposal needing approval | Brownfield and legacy, where the codebase outweighs the spec |
| Kiroprimary | Full IDE from AWS; popularised EARS requirements notation | AWS-centred teams wanting requirements rigour |
Coding-agent harnesses
By 2026 the leading tools stopped being CLIs that call a model and became harnesses in Böckeler’s sense: persistent memory, hooks, skills, isolated subagents, parallel worktreesprimary. Thoughtworks moved Claude Code to Adopt on its Technology Radarprimary.
- Claude Code — skills, subagents (which cannot spawn sub-subagents), git worktrees; practitioners report 3–5 concurrent subagents and 4–8 worktrees per developerprimary
- OpenAI Codex — cloud sandbox agent returning finished PRs; CLI and desktop appsprimary
- Google Antigravity 2.0 — agent-first IDE, desktop app, Go CLI, SDK, Managed Agents API; 76.2% on SWE-bench Verifiedprimary
- OpenHands — the leading open-source platform; CodeAct paradigm, sandboxed Docker, a proprietary critic modelprimary
- SWE-agent / mini-swe-agent — the research scaffold; the mini version is ~100 lines of Python and scores above 74% on SWE-bench Verifiedprimary
That last entry is the most instructive line on this page. A hundred-line scaffold within a few points of elaborate commercial systems is a standing argument that complexity in the harness is not the same as capability.
Commercial autonomous agents
| Company | Valuation | Reported |
|---|---|---|
| Cognition (Devin) | $26B postreputable | ARR $37M → $492M in twelve monthsreputable; best on well-bounded tasks, weak on ambiguous requirements |
| Factory.ai (Droids) | $1.5B postreputable | 58.75% on Terminal-Benchvendor-reported; “31× faster delivery”vendor-reported |
Two companies with similar names that are routinely confused: Cognition builds Devin, Factory.ai builds Droids, and they are unrelatedreputable. Note the tier chips in the right-hand column: the valuations are reported by press at fundraise, the capability and speed claims are the vendors’ own.
Orchestration
LangGraph, CrewAI and AutoGen/AG2 are the three that survived into production usereputable. LangGraph is the most production-ready: checkpointing, audit trails and rollback all fall out of having an explicit graph, which is chapter 7’s argument in library form. CrewAI prototypes fastest. AutoGen has the strongest code execution.
Benchmarks, and why to discount them
SWE-bench is foundational and saturated: 2,294 real GitHub issues across 12 Python repositories, with Verified as OpenAI’s human-filtered subsetprimary. The mid-2026 leaderboard runs 80 to 94%reputable, and OpenAI itself published a note in February 2026 on why Verified no longer measures frontier coding capabilityprimary.
Successors exist because the field noticed: Terminal-Bench for command-line tasksprimary, SWE-bench-Live for novel issues, and a family of critical benchmarks designed to expose inflationprimary.
The line to carry into a procurement conversation: none of these scores is a mergeable pull request.
Beyond code
The same harness pattern now runs content and documentation pipelines, and the gate placement is identical. Sanity’s Content Agent stages every change as a draft; nothing publishes without a human approvingprimary. a16z reports the enterprise sweet spot at 3 to 7 agents per workflow (researcher, writer, critic, publisher), with human review at write and side-effect boundariesreputable.
The design principle generalises cleanly: broad reads, narrow typed interruptible writes, human approval at the point of irreversible action. The site you are reading is an instance of it; see the case studies.
One naming caution: Microsoft’s Agent Factory is a parallel coinage for its enterprise agent-platform blueprint, not the software factory this site describesprimary.