Engineering at AI speed — Redesigning contracts, team structures, and QA for 3x faster custom development in 2026

In May 2026, InfoQ released Presentation: Engineering at AI Speed: Lessons from the First Agentically Accelerated Software Project, and around the same time OpenAI unveiled a case study, Simplex rethinks software development with Codex, featuring "software development incorporating Codex across all phases."

What both share is redesigning the development process itself to be "agent-led" rather than merely "AI-assisted." This represents the next stage beyond "tool selection" covered in Comparing Claude Code, Codex CLI, and Copilot CLI across QCD, turning the focus toward redesigning development processes rather than tools.

In custom development, responding to customer questions like "how many times faster did it get by introducing AI?" requires redesigning the delivery structure itself. This article provides an implementation guide structured across contracts, team organization, and QA.

Why development does not accelerate with simple "AI assistance"

Concerns like "we adopted AI, but things haven't sped up as expected" are typical in custom development projects. The InfoQ presentation identifies the following three causes:

These are "process challenges" preceding the measurement issues covered in Measuring custom AI ROI with DORA, SPACE, and Core 4; without establishing a "process that accelerates" prior to measurement, the numbers will not materialize.

The Simplex approach — Five redesigns for "agent-led processes"

Redesign 1: Shift specifications into formats agents can parse

Specifications written for humans are difficult for agents to parse, leading to situations where "ambiguous phrasing increases incorrect implementations." Converting them into Gherkin / BDD acceptance criteria significantly boosts an agent's first-attempt success rate.

Redesign 2: Break PRs down from "feature units" to "change units"

A "1 PR = 1 feature" structure designed for human review is too large for AI-driven workflows. By breaking them down into "1 PR = 1 change (dozens of lines)," review times drop and development cycles transform.

Redesign 3: Mandate a test-first approach in contracts

Agents can "verify on their own if tests exist beforehand," but when humans write tests afterward, operations cannot keep up with AI speeds. Include a clause that "test code must be written as a deliverable prior to implementation."

Redesign 4: Redefine "reviewers" as "pair-programming supervisors"

Shift PR reviews from "post-hoc quality gates" to "active conversational partners during implementation." Reviewers maintain real-time awareness of "what the agent is building" and issue course-correcting instructions.

Redesign 5: Redefine the definition of done from "it works" to "backed by evidence"

Simply stating "it works" is insufficient; the definition of done requires a three-part evidence package consisting of "passing tests, benchmarks within thresholds, and clean security scans." This aligns with the "declarative QA" philosophy discussed in Declarative security for custom development with GitHub CodeQL.

"AI speed clauses" to include in client contracts

In particular, "AI output liability" is a crucial clause that explicitly clarifies that stating "we are not responsible because AI wrote it" is unacceptable.

Team organization design — Four roles supporting "3x speed"

[案件オーナー] : 顧客との合意形成・スコープ管理
       │
       ├──→ [テックリード（人間）] : アーキ設計・難所対応・最終判断
       │
       ├──→ [監督者（人間）]      : エージェントの方向修正・レビュー
       │
       └──→ [Agent Fleet]         : 並列で実装（Codex / Claude Code）
              ├ Implementer Agent
              ├ Test Generator Agent
              └ Documenter Agent

By operating an implementation team that previously required 5 to 8 people with "2 to 3 humans + an Agent Fleet," teams can target 3x the output without cutting headcount. This represents a compact unit of the "company-wide rollouts" discussed in Custom AI development with a centralized multi-repo map.

QA design — Four layers ensuring speed does not kill quality

Keeping L4 strictly human is an ironclad rule to prevent the self-referential loop of "AI approving AI."

Four common pitfalls

Pitfall 1: PR reviews remaining a bottleneck

Even if PRs are broken down into smaller pieces, speed will not improve if reviewers spend their usual 30 minutes per review. Shift workflows so that the reviewer role becomes a "supervisor" providing real-time course corrections.

Pitfall 2: Blind faith in test generation AI

Test-generating agents frequently fail by "writing tests that replicate the same mistakes as the implementation." Stipulate in the contract that at least one test case must be handwritten by a human.

Pitfall 3: Delayed synchronization with clients

Moving quickly under agent-led workflows can create situations where client decision-making cannot keep pace. Maintain alignment through twice-weekly short checkpoints.

Pitfall 4: Degradation of handover quality

Agent-written code can be "difficult for humans to read," leaving client engineers stranded during the maintenance phase. Stipulate linters and formatters aligned with "human-readable" standards in the contract.

Summary — Moving from "AI-assisted" to "agent-led"

What Engineering at AI Speed and the Simplex case study demonstrate is "reinventing development processes" rather than "merely introducing tools." To have clients genuinely state that "incorporating AI made things 3x faster" in custom development, contracts, team structure, and QA must all be redesigned simultaneously.

How far you can take an agent-led delivery setup depends heavily on your existing workflows, contracts, and team makeup. For inquiries such as "we introduced AI but don't feel any faster" or "we want to ship a new product at 3x speed," we provide customized estimates after assessing your project's specific context. Please feel free to reach out via our inquiry form.

Sources

• Presentation: Engineering at AI Speed（InfoQ）
• Simplex rethinks software development with Codex（OpenAI）
• Comparing Claude Code, Codex CLI, and Copilot CLI across QCD (GH Media)
• Measuring custom AI ROI with DORA, SPACE, and Core 4 (GH Media)
• Declarative security for custom development with GitHub CodeQL (GH Media)
• Custom AI development with a centralized multi-repo map (GH Media)