State of AI in Platform Engineering Vol. 2¶
- Source: State of AI in Platform Engineering, Volume 2 (https://weaveintelligence.io/research/state-of-ai-in-platform-engineering-2026) (Weave Intelligence, 2026; authors Sam Barlien, Luca Galante, Florian Lipp; partners Thoughtworks, Vultr, Syntasso, env zero, Microsoft). Methodology (p. 55): 350 global practitioners, data collected from April to August 2026, 24 questions; profile: individual contributors 31%, team leads 27%, heads of function 19%, VP/director/C-level 23%. The maturity levels and the gains are self-reported (p. 18, p. 21). Page numbering below = the PDF's printed numbering.
- Checked on 2026-10-29: the links open.
Summary¶
- The report's central finding: 38% of organizations ship at least twice as much since adopting AI, but only 8% describe the return as transformative. The whole report explores this gap (p. 8, p. 54).
- The bottleneck is not the model, it is the platform: "platform readiness" leads the obstacles to scaling AI (27%), "model capability" is fourth (11%) (p. 24).
- Maturity pays off in steps, not in a gradient: adding agents without rebuilding the platform adds 6 percentage points of throughput; rebuilding the platform adds 24 (p. 21).
- Code generation has industrialized, absorption has not: 35% still validate AI-generated changes by human reading alone and only 17% run automated validation loops (p. 26).
- Governance is the weakest point: a quarter run agents with borrowed human credentials, 12% with a shared service account and 18% do not know how agent identity is managed (p. 38). The Agentic Engineering Platform (AEP) is presented as the evolution of the IDP, not its replacement (p. 28-30).
Numbers¶
| Data point | Value | Page | Suggested use in the presentation (act/theme) |
|---|---|---|---|
| Organizations shipping at least 2x more since AI (Figure 7, "Change in engineering throughput since adopting AI": 29% with 2x-3x + 9% with 5x or more) | 38% | 8, 20, 23 | Why platforms: opening, paired with the next row |
| Organizations describing the AI return as transformative (Figure 9, "Current ROI of AI across the SDLC"; 29% pre-value, 14% negative) | 8% | 8, 22 | Why platforms: the throughput vs. ROI gap |
| Platform practitioners who used AI every day (Vol. 1) | 89% | 12 | AI adoption in platform engineering |
| Organizations with little or no return from AI (Vol. 1) | 95% | 12 | Historical context of the "AI implementation plateau" |
| Organizations at Level 1 or below 12 months ago | more than 95% | 12, 19 | Maturity evolution in one year |
| Platform delivers central LLM connectivity | 45% | 13 | Agents/AI: the platform team takes on AI |
| Platform delivers retrieval/chat interfaces | 35% | 13 | Agents/AI |
| Platform delivers agent runtimes | 27% | 13 | Agents/AI: agent runtime as a platform service |
| Platform delivers inference workloads | 27% | 13 | Agents/AI |
| Organizations with no AI mandate | 15% | 13 | AI adoption |
| AI as a company-wide productivity initiative / C-level strategic | 36% / 33% | 14 | Adoption: the mandate comes from above |
| Grassroots adoption by individual engineers / no activity | 6% / 3% | 14 | AI adoption |
| Use cases: code generation / documentation | 84% / 79% | 15 | AI adoption in the SDLC |
| Use cases: error and log analysis / IaC generation | 66% / 65% | 15 | Resource Plane and Observability: AI-generated IaC |
| LLM consumption: direct provider API / aggregated hosting / open-source self-host / sovereign provider | 52% / 26% / 11% / 3% | 16 | Security Plane: governance of access to models |
| Fear data exposure through public AI APIs | 64% | 16, 40 | Security Plane |
| Maturity: Level 1 / Level 2 / Level 3 / Level 4 / does not know | 41% / 31% / 14% / 4% / 9% | 19 | Agents operating platforms: where the industry is |
| Organizations above Level 2 | 18% | 19 | Agents/AI: few have moved past Level 2 |
| Level 4: organizations with fewer than 100 engineers | 13 of 14 | 20 | Governance: autonomy is still a small-team achievement |
| Throughput: marginal gain (<20%) / no measurable change / 2x to 3x / close to 5x or more | 39% / 23% / 29% / 9% | 20 | Measuring success: same technology, mixed results |
| At least doubled throughput by level (Level 1 / 2 / 3 / 4; Figure 8, "Share reporting at least doubled throughput (2x to 3x, or 5x and above), by agentic maturity level") | 30% / 36% / 60% / 79% | 21 | Measuring: maturity pays off in steps (n = 143 / 110 / 50 / 14) |
| Gain from agents without rebuilding the platform (Level 1 to 2) vs. rebuilding (Level 2 to 3) | 6 percentage points vs. 24 | 21 | Why platforms: the main argument |
| ROI with transformative impact by level (Level 1 / 2 / Level 3 or 4) | 2% / 9% / more than 20% | 23 | Measuring success |
| Pre-value ROI (experimentation only) / negative ROI | 29% / 14% | 22 | Measuring: cost of licenses, tokens and review |
| Obstacles to scaling AI: platform readiness / token and infrastructure cost / review bandwidth / model capability | 27% / 23% / 16% / 11% | 7, 24 | Why platforms: "nobody is waiting for a better model" |
| AI-generated code improved stability / more noise and CI failures / no impact / does not know | 39% / 12% / 29% / 20% | 25 | Integration & Delivery: quality is not the problem |
| Validation of agent changes: human reading only / automated validation loops / does not know | 35% / 17% / 13% | 26 | Integration & Delivery: the absorption bottleneck |
| Review bandwidth as the biggest bottleneck: Level 1 to Level 2 | 12% to 23% | 33 | Integration & Delivery: parallel agents hit the review wall |
| Do not track token spend / do not know whether anyone tracks it | 13% / 14% | 35 | Observability: token FinOps |
| Who tracks: centralized billing / rate limits or quotas / cost distributed per team | 35% / 22% / 17% | 35 | Observability and governance: FinOps |
| Level 1 and 2 that do not track token spend | 15% | 36 | Observability: cost as a precondition of Level 3 |
| Security posture: governs through AI-specific policies and controls / blocks or restricts | 37% / 13% | 37 | Security Plane: guardrails instead of veto |
| Agents with their own scoped identity / identity managed dynamically by policy | fewer than 1 in 3 / 19% | 38 | Security Plane: agent identity |
| Agents with human credentials / shared service account / unknown (Figure 16, "How AI identities are managed in the platform"; fewer than 1 in 3 with their own scoped identity, 19% dynamic by policy) | 25% / 12% / 18% | 38 | Security Plane: blast radius |
| Use AI to generate infrastructure as code | 65% | 39 | Resource Plane and Security: controls stop at the code boundary |
| Agents bypassing existing controls as a relevant threat | 43% | 39 | Security Plane |
| AI risks: hallucinations in production / unmanaged tool use / data residency and auditability / prompt injection | 52% / 47% / 43% (each) / 37% | 40 | Security Plane |
| Shadow AI: minimal / significant / no visibility | 47% / 27% / 8% | 40 | Security and Observability: "observability has to precede restriction" |
| Platform customers: software teams / data science and ML / business users / external / agents | 79% / 49% / 48% / 20% / 29% | 43 | Developer Control Plane: agents as platform customers |
| Expectation of the platform engineer's role: governance and guardrails / platforms for agents / no change / automated | 43% / 36% / 11% / 9% | 44 | Developer Control Plane and governance |
| Platform and data teams: separate silos / converged / converging | 28% / 6% / 10% | 45 | Portal/catalog: expansion of the mandate |
| Investment priority in 12 months: SDLC automation / agentic coding / incident remediation and security and compliance (each) / DevEx and copilots | 31% / 23% / 16% / 13% | 53 | Agents operating platforms |
| The report's forecast: organizations at Level 3 or above next year | more than 30% (today 18%) | 53 | Roadmap: where the industry is going |
Quotes¶
- "The Agentic Engineering Platform is the evolution of the IDP, not the replacement of it." (Kaspar von Grünberg, p. 28)
- "an agent without specially designed controls and validation checkpoints does not fail gracefully; instead, it generates chaos at machine speed." (p. 28)
- "Nobody is waiting for a better model." (p. 7; reinforced on p. 54)
- "Generation industrialized. Absorption did not." (p. 7)
- "An agent that cannot be identified cannot be audited, scoped, or switched off without switching off a person" (p. 7)
- "tighter governance actually raises the autonomy ceiling instead of lowering it." (p. 41)
- "Organizations that throw agents at a Level 1 substrate will find their agents moving faster than their guardrails." (p. 30)
- "AI has made code cheap and judgment expensive." (Irina Gontcharova, Microsoft, p. 26)
- "It's because they have the platform around the model." (Rickey Zachary, Thoughtworks, p. 24)
- "the discipline exists because recurring work in software development, when done by humans, was being relentlessly ad hoc ... Define the path once, govern it, and let everyone run it." (p. 16, abridged excerpt: "Define the path once, govern it, and let everyone run it")
Concepts and frameworks¶
- Four levels of agentic development (p. 18): Level 0 Human is the loop (no agents; the IDP baseline), Level 1 Human in the loop (AI suggests, human executes), Level 2 Human on the loop (agents generate PRs in parallel), Level 3 Humans as orchestrators (agents work continuously from observed signals, review by exception), Level 4 Fully autonomous (systems of agents inside guardrails designed by humans).
- Agentic Engineering Platform (AEP) (p. 27-30, glossary p. 56): what the IDP becomes when agents are its users; an evolution of the IDP, not a replacement.
- Three layers of the AEP (p. 29): tooling layer (the five planes: developer control, integration and delivery, resource, security, observability; marked as the IDP), paths layer (repeatable, governed, machine-executable routes) and agent infrastructure layer (identity, context, capability, execution, evaluation, security, observability, all as code).
- Path types (p. 29): Deterministic (pipeline, same output for the same input), Probabilistic (agent, "Evals become the test suite"), Hybrid (probabilistic generation with deterministic gates in a loop). Recommendation: name the type of each path and require deterministic gates on irreversible effects.
- The four prerequisites of Level 3 (p. 10, p. 50): dispatch path, validation loop, scoped identity, sandboxed environments, which only pay off together.
- Validation gate vs. validation loop (p. 30, glossary p. 58): a gate is a binary checkpoint; a loop returns the failure to the agent with the reason until the deterministic checks pass ("first-pass failure is convergence, not defect").
- Six return vectors (p. 34): delivery velocity, outcome predictability, operational resiliency, maintenance overhead, organizational learning, structural incentives.
- Metric "cost per accepted output" instead of cost per token (p. 36, p. 51); four functions of token FinOps: attribute every call, watch spend velocity, enforce hard limits in the platform, produce an audit trail (p. 36).
- Six coverage questions (p. 50): is there a governed, machine-executable path in Plan, Code, Review, Test, Deploy, Monitor? 0-2 yes = the substrate is the work; 3-4 = "expensive middle"; 5-6 = the constraint has already moved to cost and architecture.
- Five recommendations (p. 50-51): build the four Level 3 prerequisites together; close the validation loop before adding agent capability; fix attribution (no borrowed human credentials); gate by consequence, not by category (possible damage and speed of detection); measure value, not output.
- Governance patterns of the leaders (p. 41): policy, identity and network inherited from the workspace; human/agent privilege separation (least privilege, writes only through pull request); ephemeral workspaces; controlled, recorded failure instead of a silent incident.
- AI implementation plateau (p. 12, glossary p. 56) and "Platform as a product, golden paths, standardization" as differentiators of those who got past the plateau (p. 12).
- Thinking in Platforms and the "Paths to Outcome" model (p. 47); IDP defined in the glossary as a platform that standardizes paths, enables self-service and reduces cognitive load (p. 57).
Relevance by plane¶
- Developer Control Plane: agents are already platform customers in 29% of organizations and 43% expect the role to shift to governance and guardrails (rows "Platform customers" and "Expectation of the role"); the dispatch path as the first new path of Level 2 (p. 56); maturity in steps (the throughput-by-level rows and 6 vs. 24 p.p.).
- Integration & Delivery Plane: absorption bottleneck (35% human reading only vs. 17% loops; review bandwidth from 12% to 23% from Level 1 to 2); validation loop as the engine (p. 30); AI code does not degrade stability (39% improves, 12% worsens); Plan/Code/Review/Test/Deploy/Monitor coverage (p. 50).
- Resource Plane: 65% use AI to generate IaC and 66% for error and log analysis (p. 15, p. 39); current controls inspect code, not the effect on the provisioned environment; policy-as-code and sandboxed execution are still a minority (p. 39); gate by consequence (p. 51).
- Observability Plane: token FinOps (13% do not track, 14% do not know, 15% in Levels 1 and 2; attribution by team, path and model; "cost per accepted output"); shadow AI (47% minimal, 27% significant, 8% no visibility; "observability has to precede restriction", p. 41); 66% use AI in log analysis.
- Security Plane: agent identity (25% human credentials, 12% service account, 18% do not know, 19% dynamic by policy); 37% govern by policy vs. 13% block; risks (64% data exposure, 52% hallucinations, 47% unmanaged use, 43% control bypass, 37% prompt injection); leaders' patterns (workspace-level, least privilege, ephemeral workspaces, p. 41).
- Portal/catalog: the report does not address portals directly (no mention of Backstage or a catalog in the pages read); use only the concept of a path as a "repeatable route from intent to usable output" (p. 29), golden paths as a differentiator (p. 12) and the expansion of customers (business users 48%, data/ML 49%, p. 43) to justify a single portal for humans and agents. Kratix appears only as a product of the partner Syntasso (p. 5-6), not as a research finding.
-
Agents/AI: maturity (41/31/14/4/9%), throughput by level (30/36/60/79%), 6 vs. 24 p.p., the four Level 3 prerequisites, path types, tooling vs. agent infrastructure, the forecast of more than 30% at Level 3 or above, investment priorities (31% SDLC automation); why "platform readiness" (27%) beats "model capability" (11%).
-
Definition of the levels (Figure 5, p. 18): Level 0 "Human is the loop" (no agents; the IDP baseline); Level 1 "Human in the loop" (AI suggests, humans execute); Level 2 "Human on the loop" (agents generate PRs in parallel; humans trigger and verify behavior); Level 3 "Humans as orchestrators" (agents work continuously from observed signals; review by exception); Level 4 "Fully autonomous" (systems of agents start and finish work inside guardrails designed by humans).