Research · Area 06 of 07
Agentic-Driven Development
AI can write code; the open question is what evidence justifies trusting the result. The field’s constraint moved from generation to verification: reviewers spend far more time on AI-written code, brownfield systems resist agents that thrive in green fields, and most agent pilots never ship.
Stoa Labs studies verification-first development loops, the context engineering that makes agents useful in legacy codebases, right-sized specification, and the authority coding agents should receive in consequential repositories.
This is also where we dogfood most directly: our own operations run on the same agentic methods we research.
01 · Definition
A spectrum defined by verification
Agentic-driven development is using AI coding agents for production-focused software engineering and managing the entire software development lifecycle with agents. The field stabilized in 2026 around a spectrum defined not by tooling but by verification: who owns correctness.
Working terminology as of mid-2026:
- Vibe coding. A posture: trust the model, do not read the diff, keep prompting until it works. Fine for prototypes; collapses in production because it skips design, review, and testing.
- Agentic engineering. An architecture and a discipline: the model is wired into tools, runs in loops, and produces code a human verifies against a definition of done. The default is orchestrating agents rather than typing code, with expertise concentrated in oversight.
- AI-native SDLC. The whole lifecycle reframed around agents. The consensus insight: the bottleneck moved from typing speed to how well the problem is specified and how rigorously the output is verified. The work that bookends implementation stays human-paced.
- Spec-driven development (SDD). Treating a written spec as the primary artifact, with code as a regenerable output. A more structured branch of the same movement, currently in its first backlash.
This pillar is also where Stoa Labs dogfoods most directly: we run our own engineering on the same agentic methods we research and use.
02 · Why it matters
Where the constraint moved
- Production practice is real and accelerating. Frontier labs report large majorities of their code as agent-written, background and scheduled coding agents went mainstream in H1 2026, and multi-agent workflows over isolated git worktrees became a standard team pattern.
- But the constraint moved. Reviewers spend 91 percent more time on AI-generated code, which carries measurably more readability and logic issues than human-written code, and AI reviewers still produce heavy false-positive triage.
- Generation is commoditizing monthly. Verification, context, and methodology are where durable value sits.
03 · Open questions
What we are trying to find out
Exploration-mode questions: pursued to learn something, allowed to not pan out.
- The verification loop. What combination of reproduction harnesses, agent-written tests, small-diff enforcement, and CI gating produces trustworthy review at agent speed? Nobody has a scalable answer.
- The brownfield wall. Experienced developers in million-line repos measured 19 percent slower with AI tools while feeling faster. What context engineering (semantic code graphs, retrieval, instruction-file conventions) actually closes the gap in legacy codebases?
- Right-sized specification. SDD’s failure mode is spec overhead exceeding the code it replaces, and identical specs producing different code. What spec depth is justified at each change size, and can regeneration be made reproducible enough to trust?
- The perception gap. Developers consistently feel faster than they measure. What instrumentation gives teams a true read on agentic productivity?
04 · Directions
What this work could become
Candidate development directions, not commitments or year-one programs.
- A verification-first methodology and toolkit. The packaged answer to the review bottleneck, from reproduction harnesses to CI gates. Sticky, codebase-specific, and aligned with the portfolio-wide verification thesis.
- A brownfield agentic-engineering practice. Context engineering for legacy systems is under-served, technically deep, and exactly where mid-market codebases live.
- A pilot-to-production playbook. Around 88 percent of agent pilots never ship; the known success predictors (named owner with budget, automated evals before deploy, single-workflow scope with binary success criteria) can be systematized into a repeatable engagement.
- Right-sized SDD. Spec discipline calibrated to change size, positioned as skeptic-practitioner rather than evangelist.
05 · Boundaries
What we deliberately avoid
Competing with the generation layer. Coding agents are cheap, excellent, and shipped by frontier labs; anything we build there is obsolete next quarter. Our IP sits in methodology, verification, and context: the layers that are codebase-specific and do not commoditize.
06 · Landscape · 2026-07
The landscape, July 2026
Claude Code leads on depth (with async remote tasks since March 2026), Codex closed much of the gap, Cursor dominates the IDE surface, and Devin’s price collapse signaled background-agent commoditization. Multi-agent development over git worktrees went mainstream (VS Code shipped it in February 2026; teams run several concurrent worktrees per developer).
The SDD wave crested: GitHub Spec Kit passed 93,000 stars, AWS Kiro is the most legitimately spec-first tool but faces lock-in and pricing backlash, and Tessl raised $125M for spec-as-source while its regeneration remains non-deterministic. Martin Fowler’s team published the definitive skeptical analysis, invoking the failed model-driven-development era.
The review bottleneck became the headline problem, and AI code reviewers (CodeRabbit, Greptile and peers) improve but do not close it. Trust lags adoption: most developers use these tools daily while a minority trust the output.
Landscape notes are dated on purpose. This layer moves monthly, and we treat any claim older than a quarter as due for reverification.
07 · Sources
Selected sources
- Simon Willison, "Vibe coding and agentic engineering" (May 2026) and the Agentic Engineering Patterns collection
- Martin Fowler / Birgitta Böckeler, "Understanding Spec-Driven Development: Kiro, spec-kit, and Tessl"
- CodeRabbit 470-PR review-time analysis; Greptile code-review benchmarks
- Anthropic 2026 Agentic Coding Trends Report; VS Code multi-agent development (February 2026)
Where this work goes
The work anchors our agentic development and pilot-to-production services.