Agent Experience: Designing Internal Developer Platforms That AI Agents Can Consume

Agent experience (AX) is the practice of designing platform capabilities so autonomous AI agents can use them correctly without human interpretation. It is what turns a conventional platform into an agent-ready internal developer platform (IDP), and it is the platform half of a workforce that now includes engineers and agents both.

The breakthrough focuses on engineers and AI agents working as a single delivery system, with engineers carrying judgment and accountability and agents carrying scale and repetition. That system runs on a platform, and the platform was designed for only one of them.

McKinsey’s State of AI 2025, found 62% of organizations at least experimenting with AI agents and 23% already scaling an agentic system somewhere in the enterprise. Inside any individual business function, no more than 10% report scaling agents at all. Breadth without depth is the signature of a population that spreads through existing access paths rather than designed ones.

Platform teams inherit that population without having designed for it. AX names the design work that closes the difference, and three questions define it: why developer experience practices do not transfer, what makes a capability agent-consumable, and how a team measures whether it is working.

Developer experience assumes the user can resolve ambiguity. Documentation fills gaps the interface leaves open. A README explains the convention the API does not enforce. A Slack channel absorbs the questions the portal fails to answer. This works because human engineers treat an unclear situation as a signal to stop and ask.

An agent treats an unclear situation as a problem to solve. It infers a convention from the closest available pattern, proceeds, and applies the same inference across every subsequent invocation. A gap that produced one confused developer now produces systematic deviation across every repository the agent touches.

AX work therefore moves effort out of documentation and into interface design. Every convention an agent needs has to exist inside the contract the agent can read at execution time.

• Discoverable without documentation. An agent should be able to enumerate what a platform offers, and understand what each capability does, by querying the platform itself. A catalog that lives on a wiki page is not discoverable. This is why the Model Context Protocol and similar tool-description standards have gained enterprise traction: they give agents a consistent way to find out what is available before acting.
• Deterministic in its response contract. Anthropic’s engineering team describes a tool as a contract between a deterministic system and a non-deterministic one, which is different from the API-to-API guarantees platform teams are used to writing. A capability that returns different shapes under different conditions teaches inconsistent lessons, and the inconsistency propagates.
• Explicit about failure. The most valuable AX improvement available to most platform teams is better error responses. Anthropic’s guidance makes the same point from the model side: error responses should communicate specific, actionable corrections rather than opaque codes or tracebacks. An error that names the violated constraint and the acceptable alternative turns a failure into instruction.
• Bounded in blast radius. Agent-consumable does not mean agent-permitted. Every capability needs a scope limiting what a single invocation can affect, independent of what the agent was asked to do. Rate limits, resource quotas, and environment boundaries do for agents what code review does for humans.

Platform teams already measure developer experience through adoption, lead time, and satisfaction. Agents do not fill in surveys and or issue complaints, so AX needs its own instrumentation, and building it is the practical starting point for AI platform engineering.

Four measures give teams a usable picture.

1. Improvisation rate counts actions taken outside the paved path.
2. Retry ratio counts repeated attempts against the same capability, which usually indicates unclear error responses rather than transient failure.
3. Catalog coverage measures the proportion of capabilities exposed in machine-readable form.
4. Scope breach attempts count actions blocked by permission boundaries, showing where agent intent and platform policy disagree.

Anthropic arrives at a similar set from the model side, recommending teams track tool call counts, runtimes, token consumption, and errors, and read them diagnostically: redundant calls point to pagination limits that need resizing, and repeated parameter errors point to descriptions that need clarifying rather than agents that need retraining.

The reason to measure is that the risk is currently unmanaged. McKinsey found 51% of organizations using AI have already seen at least one negative consequence, and that explainability, the second most reported risk, is not among the risks most commonly mitigated.

AAVA™, Ascendion’s agentic AI platform for enterprise software engineering, is model agnostic and supports Model Context Protocol, Agent-to-Agent, and Agent Communication Protocol standards, so agents can discover and call capability through interfaces built for them. It ships with 4,000+ pre-built engineering agents and 2,500+ ready-to-use workflows, and integrates with the systems engineers already use: Jira, GitHub, Confluence, and ServiceNow. Certified “Golden Agents” are published centrally and shared across teams, so an agent or its operator can enumerate what has already been validated rather than rebuilding it.

The sharpest test of agent-consumable design is what agents do with a system that offers them nothing. On a 40-year-old platform at a digital-first bank, with no available documentation, developers, or subject matter experts, Ascendion used AAVA to reverse-engineer 900,000+ lines of 1980s code in three weeks, then generate 4,200+ requirement use cases and 23 go-to-market capabilities mapped against 60 business capabilities. With legacy modernization services, Ascendion completed the project in half the time, and at a third of the cost of traditional methods.