Someone will eventually ask whether your internal developer platform needs rebuilding for the agent era. The instinct is to treat build-versus-buy as the whole decision, before anyone has defined “ready” or modelled the return. Agent requirements shift monthly, and shadow AI inflates the price of doing nothing.
Run the rebuild decision building on what changes in Platform Engineering 2.0: assess readiness, cost the ROI, pick a path the board will fund.
How do you evaluate whether your internal developer platform is ready for AI agents?
Evaluate your platform against four agent-native capabilities. Can it issue per-agent, session-bound machine identities? Enforce policy at machine speed, with rate limits and least privilege? Reconstruct agent behaviour from correlation IDs and structured logs? Isolate workloads so a misdirected agent cannot reach production?
Machine identity is the foundation. The test moves from “does this agent have the right key?” to “was this agent authorised for this action, in this context, at this moment?” A static credential cannot answer that (ReadySetCloud). A modernisation review measures architectural maturity; agentic readiness measures whether the platform can safely handle machine-speed interaction. A platform can pass one and fail the other. Shadow AI is your leading readiness signal. See agent-native readiness and machine identity.
So what can you safely hand over?
Which platform capabilities are safe to expose to AI agents, and which need hardening first?
Readiness is a per-capability score. Score each capability across the five pillars: interface readiness, security and identity, data handling, operational resilience, and observability. Then apply severity ratings (blocker, risk, advisory) by service archetype. A gap is advisory when an agent only queries a stateless utility or a data store, but a blocker when it modifies state through a CRUD service or an orchestrator (AWS).
Expose read-only, low-blast-radius surfaces first: service-catalogue queries, golden-path discovery, template scaffolding. Harden anything that writes, deletes, changes state or spends money first: provisioning, deploys, secrets, model selection. A model registry answers “which models may agents use?” before any capability is wired in (Red Hat). Gate with least privilege and per-action authorisation, not broad roles. Databricks puts it well: you control what an agent can access and monitor what it actually does (Databricks).
The next question is cost.
How do you assess the ROI of evolving your platform for the agent era?
Anchor the case to a real number. An enterprise internal developer platform is typically run by a ~60-person product organisation and costs roughly $7.5 million a year, per The Next Platform, cited by Develeap.
ROI is business value minus total cost, divided by total cost, spanning build, inference, integration and maintenance. FinOps keeps spend legible with cost per inference, token growth and budget guardrails (FinOps Foundation). A model registry keeps model costs visible. Fold in shadow AI and validate with delivery metrics. Accenture and the WEF titled their review Proof over Promise; it found measured AI value routinely trails the promise.
First, the full TCO picture.
What does the total cost of ownership of AI agents include beyond licensing and inference?
Beyond licensing and inference, agent TCO includes integration, data preparation, maintenance and compliance. Data preparation eats 60 to 80 per cent of effort, post-deployment work can be around 65 per cent of lifecycle cost, maintenance adds 15 to 25 per cent of the build, and most organisations misestimate AI cost by more than 10 per cent, with nearly a quarter underestimating by 50 per cent or more (Keyhole Software).
Model requirements move monthly, so maintenance is a standing cost. One vendor logged three breaking changes in a month across Claude Code, the Agent SDK and Claude Cowork (Anthropic’s April 2026 postmortem). Ignore that churn and the rebuild case understates cost and overstates return, the verification tax in the trust-gap metrics.
Much of that cost is already shadow AI.
How do you measure the cost and risk of shadow AI across your organisation?
Measure shadow AI in three buckets: unauthorised spend, breach exposure, and rework after audits. Pull licensing, API tokens and inference from SSO, expense and FinOps telemetry. IBM’s 2025 benchmark puts shadow AI at roughly $670,000 added to the cost of a breach, and shadow AI features in roughly one in five incidents (Reco). Count vibe-coded tools as the observable tip of the iceberg.
Aggregate it into one price-of-inaction figure. Do not just shut shadow AI down; that drives it underground. Make the sanctioned platform the path of least resistance (Syntasso).
With that priced, build-versus-buy gets clearer.
What should a CTO look for when deciding whether to build, buy or compose an agent-ready platform?
The decision has three options: build, buy or compose. Score each against four criteria: requirement velocity, in-house capacity, identity and security depth, and FinOps tooling maturity (VMware). Because requirements shift monthly, fixed purchases and bespoke builds both break.
Compose is usually where the economics land: buy the commodity (foundation models, embeddings, developer tools) and build the integration layer you own. Blake Romano of Imagine Learning notes the agent SDK will keep changing, so your organisational context is where the durable value sits (Buoyant). Cross-check the cost benchmark (The Next Platform, via Develeap) and the spend controls (FinOps Foundation).
Whichever path you pick, the numbers decide.
Activity metrics vs delivery metrics: which ones actually reveal agent ROI?
Adoption, PR volume and generated lines are activity theatre; they can climb while value stays flat. Delivery metrics reveal whether agent spend shipped anything: feature completion rate, change failure rate, lead time and review-queue health.
AI tool usage rose 65 per cent over 15 months while PR throughput moved 8 per cent (GetDX). METR found developers took 19 per cent longer while believing they were 20 per cent faster (METR). Faros AI saw PR volume jump 98 per cent per developer with flat DORA metrics, while median review time rose 91 per cent (Augment Code). Pair throughput gains with change failure rate to catch false velocity. That is the DORA data behind the trust-gap article.
The agent-era platform decision rests on three questions: are we ready, what will it really cost, and which path is cheapest to change when requirements move again? Answer those in order and hold the result to delivery outcomes. Then the rebuild call becomes a defensible business case. See the complete Platform Engineering 2.0 overview.
Frequently Asked Questions
Do we actually need to rebuild the whole platform for the agent era?
No. Rebuilding is the last question, not the first. Start by assessing readiness against the four agent-native capabilities, then expose low-blast-radius capabilities and let delivery metrics tell you whether a bigger investment is justified. Most teams find the path is to harden identity and observability, not rip out the platform. Treat “rebuild” as the outcome of a costed business case, never the opening move.
What is the difference between machine identity and an API key?
An API key is a static credential, so it is a shared secret that can be copied, leaked and reused. Machine identity is contextual. It binds an agent to a specific action, in a specific context, at a specific moment, usually through session-bound credentials. That is the difference between “has the right key” and “is authorised for this action right now”, and it is the load-bearing pillar of agentic readiness.
What should we do if we discover shadow AI is already in use?
Do not just shut it down, that drives it further underground. Measure it first. Pull unauthorised tool spend from SSO, expense and FinOps telemetry, count the vibe-coded tools, then aggregate a price-of-inaction figure. Feed that number into your readiness signal and your ROI denominator, and use it to make the sanctioned platform the path of least resistance.
How often should we reassess platform readiness?
Reassess readiness every time agent requirements move, which in practice means monthly or quarterly rather than annually. Model capabilities, especially Anthropic Claude, shift monthly, and a capability that was safe to expose last quarter may need hardening this quarter. Bake the reassessment into your delivery cadence so the rubric tracks reality instead of going stale.
Is it true that AI will immediately cut delivery costs?
Not automatically. Headline throughput gains are often eroded by verification and review cost, which is why activity metrics can climb while value stays flat. Treat the AI verification tax as part of your ROI denominator and validate spend with delivery metrics like change failure rate and lead time. Cost reduction only shows up when stability holds and features actually complete, not when lines of code grow.
What is the difference between a modernisation assessment and an agentic readiness assessment?
Modernisation measures architectural maturity, how clean, scalable and maintainable the platform is. Agentic readiness measures something narrower and sharper: whether the platform can safely handle machine-speed interaction through machine identity, policy, observability and workload isolation. The two are complementary, not interchangeable. You can have a modern platform that is still not safe to hand to agents.
What does “compose” actually mean when building an agent-ready platform?
Compose means buying commodity and building only the integration layer that is genuinely yours. You purchase foundation models, embeddings and developer tools, then write the domain-specific wiring that connects them to your catalogue, golden paths and policies. It sits between buying everything and building everything, and it keeps the replaceable parts replaceable while you own the actual differentiator.
Which metric should we watch first when agents start shipping code?
Pair change failure rate with whatever throughput metric you are celebrating. Throughput alone produces false velocity, as Faros AI’s 98% PR-volume jump with flat DORA metrics shows. If change failure rate rises while PR volume climbs, you are trading speed for instability. Watch that pair first so you catch the verification tax before it becomes review-queue debt.
What counts as shadow AI?
Shadow AI is any AI or agent use running outside sanctioned platform paths. That includes unauthorised ChatGPT or Claude subscriptions on the corporate card, API tokens bought outside procurement, vibe-coded internal tools, and agents that bypass identity and policy controls. If it is not governed, observable and gated by the platform, treat it as shadow AI regardless of how harmless it looks.
Do we need FinOps in place before we start using agents?
Not as a full function, but you need the telemetry. FinOps is the control layer that makes cost per inference, token growth and budget guardrails legible, and without it your ROI denominator is guesswork. Start with the three signals that matter: cost per inference, token usage and request-volume growth. Add budget guardrails early so agent spend is capped before a cheap pilot turns into an expensive platform.
