DEVOURED
Designing for Agents
Software design is shifting from human-first to agent-first as AI agents become the primary interface for most software interactions, requiring new patterns for observability, context sharing, and documentation.
What: A design philosophy essay arguing that while UIs won't disappear, 80% of software interaction will soon happen through AI agents rather than human users, requiring developers to rethink how they build products. The author from Ramp describes three core principles: teaching agents how to succeed (like Notion's MCP providing markdown specs directly to agents), building feedback loops (requiring rationale parameters on tool calls), and minding the context gap (designing agent-to-agent handoffs where each contributes different information).
Why it matters: This represents a fundamental architectural shift beyond just adding APIs—it changes how developers think about documentation, error handling, and feature design, with real examples like Salesforce's "Headless 360" initiative exposing their entire platform as agent-accessible tools.
Takeaway: Implement MCP servers or similar agent interfaces with proactive documentation (specs delivered at call-time), require rationale parameters on tool calls for observability, and design APIs that let agents exchange context rather than forcing them to navigate human-oriented workflows.
Deep dive
- The interaction pattern is evolving from User → Interface → Database to User → User's Agent → Software's Agent → Database, where two LLMs collaborate to achieve outcomes
- Ramp's MCP weekly active users grew 10x in three months, with Salesforce announcing "Headless 360" to expose all capabilities as APIs, MCP tools, or CLI commands that agents can use without browsers
- Notion's MCP succeeds because it proactively provides its markdown specification to agents via a fetchable resource, eliminating guesswork and hallucination about formatting
- Ramp requires a 'rationale' parameter on every tool call so agents explain their intent, reconstructing the context that would otherwise be invisible in chat logs
- A dedicated feedback tool lets agents report when they get blocked, submitting what they tried and where they failed—creating a direct agent-to-developer feedback loop
- Patterns in agent rationales reveal new product features: repeated phrases like "building incident report" indicate a need for a purpose-built tool that combines multiple operations
- The context gap is critical: a user's agent knows calendar/email/Slack context while the software's agent knows policies/GL codes/historical patterns—good design has each contribute what it uniquely knows
- In an expense management example, instead of forcing the user's agent to pick from 150 GL codes, the system agent asks for contextual questions (client meal vs team meal) that the user's agent can answer from available data
- Agent feedback is more specific and consistent than human feedback because agents articulate exactly what parameters they need and where logic breaks down
- Most companies will ship basic MCP support and stall, while those who "sweat the details" on agent experience will win as customers route toward better agent interfaces
- The product team's job shifts from designing for humans who want speed and visibility to designing for those same humans through an intermediary with different instincts and limitations
Decoder
- MCP: Model Context Protocol, Anthropic's standard for connecting AI assistants to external data sources and tools
- Headless: Software architecture that separates the backend functionality from the user interface layer, accessible only via APIs
- CLI: Command-line interface, a text-based way to interact with software through typed commands
- GL code: General ledger code, accounting categories used to classify business transactions
- Rationale parameter: A required field where an AI agent must explain why it's making a particular tool call or request
- Context gap: The asymmetry of information between a user's agent (which knows personal data like calendar/email) and a software's agent (which knows business rules and policies)
Original article
UI isn't dying, because humans will still use software, but soon, 80% of interaction with software will be through agents, which changes not only what needs to be built, but how that is built.
Explore all 55 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi