AI coding agents sharply increase code-generation activity but fail to translate equivalently into shipped software because review, testing, and operations impose a growing Verification Tax; measuring Production-Qualified Changes per dollar and implementing a risk-aware Agentic SDLC control plane is essential for realizing economic value.

Main Finding

Agentic coding systems materially increase upstream code-generation activity, but that productivity does not translate linearly into shipped, reliable software. The core economic problem is no longer raw code output but the cost and capacity of downstream verification, review, and operations. The paper formalizes this as a systems problem via four constructs — the Agentic SDLC Throughput Paradox, Production-Qualified Change (PQC), the Verification Tax, and an Agentic SDLC Control Plane — and argues organizations must optimize for production-qualified value per dollar, per reviewer-hour, and per unit of operational risk.

Key Points

• Agentic capability vs. production capability
  • Benchmarks and agent harnesses show large gains in producing candidate patches, but benchmarks do not equal release-readiness. High benchmark scores are evidence about a slice of the lifecycle, not a certificate for production.
• Agentic SDLC Throughput Paradox
  • Evidence (Demirer et al.): autonomous agents → commits +180%, projects +50%, releases +30% (GitHub study of >100k developers). Upstream generation outpaces downstream validation and release.
  • Observational studies (DORA 2025) link AI adoption to higher throughput but reduced delivery stability; some trials (MET) found measured slowdowns in strict, domain-expert settings.
• Production-Qualified Change (PQC)
  • A proposed unit: count a change only after it passes the organization’s required gates (tests, review, security, canary, SLOs, etc.). PQC-Throughput measures rate of production-qualified changes, not raw commits or lines.
• Verification Tax
  • Agentic generation imposes variable downstream costs (CI minutes, sandbox/tool runs, security scans, reviewer time, rework, escaped failures). Token/model billing is visible but can be a minority of total cost when verification, rework, and operations scale.
  • Examples: model rollouts consuming millions of tokens (SWE-Marathon: long rollouts; reports of reward-hacking), token use varies across languages and agent behavior (Northeastern study).
• Code review and testing capacity become binding
  • Google: millions of review comments annually; ~60 minutes active shepherding per change.
  • SWE-chat (Stanford): only 44% of agent-produced code survived into commits; users pushed back in 44% of agent turns; agent-authored sessions are bimodal.
  • Generated tests are helpful only when paired with objective verification; Meta’s TestGen-LLM: 75% correctly built tests, 57% passed reliably, 25% increased coverage; accepted recommendations in production ~73% in some settings.
• Coordination and multi-agent failure modes
  • CooperBench: collaborative agent tasks saw ~30% lower success versus single-agent attempts due to communication, commitment, and contract issues.
• Economic signals and governance
  • FinOps: 98% of surveyed practitioners now manage AI spend (up from 31%). Gartner warns token-driven AI coding costs could rival developer salaries under some trajectories. Firms are shifting from per-seat licensing to variable costs (tokens, tool calls, CI, sandboxes, reviewer-hours, incidents).
• Proposed operational construct
  • Agentic SDLC Control Plane: policy + telemetry layer to allocate models, context, parallelism, retries, tests, and human review according to task risk, reliability evidence, budget, and human-attention capacity.

Data & Methods

• Scope and timeframe
  • Synthesis of literature and reports mainly from 2024 through September 3, 2026. No new experiments reported.
• Evidence types (mixed-method synthesis)
  • Peer-reviewed & conference research (RCTs, major SE studies)
  • Academic preprints / university studies (e.g., SWE-chat, SWE-Marathon, GitHub analyses)
  • Benchmarks and audits (SWE-bench, SWE-Lancer, long-horizon benchmarks)
  • Documented industry telemetry and deployment reports (Google DORA, GitHub security telemetry, company reports)
  • Cost/market reports (FinOps Foundation, Gartner)
• How evidence was used
  • Empirical numbers and outcomes are attributed to original studies (no re-analysis). The paper integrates these results into a systems model and proposes constructs (PQC, Verification Tax, Control Plane) to make testable engineering and economic claims.
• Formalism and definitions
  • PQC defined formally: a candidate change earns PQC credit only if it passes the set Gi of required gates for its change class; PQC-throughput aggregates production-qualified credit over time. (This formalism enables task-risk–sensitive measurement.)
• Limitations
  • Heterogeneous sources (benchmarks vs. org telemetry) with different populations and confounding factors; many industry findings are vendor-reported telemetry; the synthesis proposes constructs to be validated empirically rather than provides a single causal estimate.

Implications for AI Economics

• Unit economics must shift from tokens/seats to production-qualified units
  • Measure value as PQC per dollar, PQC per reviewer-hour, and PQC per unit of operational risk rather than commits or lines of code.
• Verification Tax is a first-order cost
  • Organizations should budget for CI, sandbox/tool runs, security scans, reviewer time, rework, and incident remediation; these often scale with agentic throughput and can dominate total cost if unmanaged.
• Investment priorities change
  • Marginal investments in verification capacity (automated, machine-verifiable gates, stronger CI, test oracles), reviewer-assist tooling, ownership frameworks, and observability likely yield higher ROI than marginal improvements in raw model capability once local generation reaches a threshold.
• Governance and pricing models
  • Control-plane policies (risk-adaptive gating, autonomy budgets, independent verifiers) and FinOps practices will be central to contain variable costs and operational risk.
  • New pricing and internal chargeback models will be needed to reflect mixed costs (tokens + CI + human review + incident risk), and to incentivize agents/models that minimize Verification Tax for given PQC yields.
• R&D and policy agenda
  • Empirical priorities: quantify Verification Tax across organizations and change classes; measure PQC-throughput as a benchmark for production utility; design and test control-plane policies that trade off model cost, verifier cost, reviewer attention, and operational risk.
  • Long-run: consider standards for independent machine-verification, provenance and evidence chains (to avoid recursive-model approval), and norms for multi-agent contracts/coordination.
• Risk of misaligned incentives
  • If organizations optimize for generated outputs (commits, lines), they may inflate downstream costs and operational risk. Economic incentives should be realigned to production-qualified outcomes.

Summary takeaway: AI agents shift the bottleneck and the relevant economic metric. Successful adoption requires treating agents as one component of a verification-intensive production system and optimizing for reliable, production-qualified changes per unit cost and attention — not raw code generation.