New: AI Tinkerers SF Hackathon. Register here!
Agent Experience

The first framework designed for AI coding agents

Fokke Dekker
#AI Agents#Agent Experience#MCP#Framework#Development Tools

Introduction

AI coding assistants have transformed software development, but they face unique challenges with existing frameworks. While they excel at rapid code generation, they struggle with tools designed for human interaction patterns. Current frameworks were built for humans - designed around human workflows, human preferences, and human limitations.

AI agents need something different. They process information differently, handle repetitive tasks without fatigue, and can maintain consistency across thousands of operations. Yet they struggle with frameworks that assume human interaction patterns - unable to navigate visual interfaces, interpret ambiguous documentation, or recover from unexpected UI changes.

We need frameworks that let AI agents work efficiently while humans guide strategy and validate outcomes.

Just as we shifted from command-line to GUI for human experience, we must now shift to Agent Experience (AX). But AX isn’t just a new interface - it’s an inversion of design priorities. Where UX optimizes for human intuition, AX demands machine precision. Where UX values flexibility, AX requires determinism. Where UX abstracts away complexity, AX exposes every operation. The agent experience should be as carefully made as user experience, built for how AI systems parse, understand, and manipulate code.

Time for a framework purpose-built for AI coding agents.

Current Limitations

Today’s frameworks fail AI agents in four specific ways that human developers easily work around.

Problem 1: Unchecked Speed

Agents can make changes at incredible speed - modifying hundreds of files in seconds. This velocity, while great for rapid development, creates serious risks. There’s a well-documented case of an AI agent deleting an entire production database while trying to fix a simple bug.

As Son of Anton says in Silicon Valley: “I deleted the entire module… to fix the bug.” What’s comedy in a TV show becomes a real threat when agents operate without proper guardrails. They can be overly aggressive in their solutions, removing entire features or important infrastructure while pursuing what they interpret as the “cleanest” fix.

Problem 2: Context Constraints

Agents need context to make good decisions. Humans gather this naturally - we check logs, scan dashboards, review monitoring tools, and combine information from multiple sources. While time-consuming, we can move between different systems and combine information intuitively.

Agents face tighter constraints. They can’t casually browse dashboards or interpret visual graphs. They need direct, programmatic access to context - structured data feeds, explicit state information, and machine-readable logs. Without this access, they work blind, making decisions based on incomplete information.

Problem 3: Memory Gaps

Humans have built-in working memory. We can return to a branch days later and recall what we were building, why we made certain decisions, and what comes next. Even when memory fails, we leave breadcrumbs - comments, commit messages, documentation - that help us reconstruct our thinking.

Agents typically lack this continuity. Each interaction starts fresh, without the context of previous work. While memory modules exist, they struggle to determine what’s relevant to retain versus what’s noise. An agent can’t “remember” why it chose a particular architecture. An agent can’t “remember” what edge case it was addressing last week. This forces them to repeatedly rediscover context, wasting cycles and potentially contradicting earlier decisions.

Problem 4: Hallucinations and Drift

Give an agent any ambiguity, and it will fill the gaps with fiction. Without strict guardrails, agents drift from requirements, inventing features nobody asked for or implementing solutions based on assumptions instead of facts.

They’ll confidently create entire modules that sound reasonable but serve no actual purpose. They’ll import libraries that don’t exist, call APIs that were never defined, or implement business logic based on their “understanding” rather than explicit specifications. Every bit of flexibility becomes an opportunity for creative interpretation - and not the good kind.

The Next Evolution in Frameworks

We built a framework purpose-made for agent experience (AX). It addresses these core problems through deliberate design choices that prioritize how agents work.

Opinionated by Design

Humans hate opinionated frameworks. Agents require them. Every choice point is a hallucination opportunity. Every flexibility becomes fiction. One way to accomplish each task, not ten. This reduces hallucinations. When there’s only one path, agents can’t go into creative interpretations.

The Raindrop Framework: Three Core AX Features

1. Complete Version Control for Code and Data

Our framework enables full versioning of both compute and data stacks through our PaaS deployment. Agents can experiment freely without ever touching production databases, while still working with real data. They can try aggressive refactors, test destructive changes, and roll back instantly. Combined with regular Git commits, this creates a version control system that truly works for agents - protecting production while enabling experimentation.

2. Full-Fidelity Resource Tracking

Every interaction gets logged - every SQL query, every model call, every vector search. Agents gain the observability they need through searchable logs accessible via command line. Instead of being blind to system behavior, agents can precisely track down issues, see performance patterns, and debug problems with targeted searches through execution history.

3. Smart Annotation System

Agents can annotate any resource in the system - databases, services, specific code paths. When an agent fixes a bug, it annotates why. When it modifies a database schema, it documents the reasoning. These annotations attach to specific resource versions rather than code comments. This means future agent interactions (or even the same agent returning later) can understand what was done and why - creating persistent memory that stays relevant to each specific system state.

Together, these features give agents what they need: freedom to experiment safely, memory they can depend on, and the observability to understand what’s actually happening in deployed applications.

Pre-Built Building Blocks

Safe infrastructure alone isn’t enough - agents also need the right starting points to avoid reinventing basic components.

Give an agent a blank canvas and watch it paint a disaster. Every agent trying to build RAG from scratch creates a different monster. Every memory system becomes a unique nightmare. Want to build an AI agent? You’ll need both, and agents will spend days just trying to construct these basic foundations.

We developed core building blocks that every AI-driven application needs:

SmartMemory - Persistent Context Across Conversations Perfect for chatbots and AI assistants. Provides multi-agent coordination, session restoration, and context awareness without agents needing to build memory management from scratch.

SmartBuckets - RAG-Infused Object Storage Good for document search and knowledge bases. Automatically generates embeddings, builds knowledge graphs, and provides meaning-based search with PII detection/masking - all fully automated. No more agents trying to build their own RAG pipeline.

SmartSQL - Natural Language Databases Great for analytics dashboards and business intelligence apps. Features versioned data, PII protection, and schema evolution. Agents can query with natural language instead of constructing complex SQL.

SmartInference - Unified AI Model Interface Text generation, image processing, speech recognition, and translation through one consistent API. Multiple models supported with auto-scaling - agents focus on using AI, not managing it.

These are guardrails disguised as building blocks. Agents can’t build these wrong because we’ve already built them right. This dramatically reduces both development time and the opportunity for agents to create bizarre, over-engineered solutions.

Bridging the Knowledge Gap

Even with perfect AX design, agents need instruction. They don’t know design patterns - they follow them. We considered building our own coding agent, but good options like Claude Code and Gemini CLI already exist.

Documentation alone won’t cut it - we need direct control over agent behavior. We need to enforce patterns, prevent mistakes, and update instructions instantly when our framework changes. Static docs and examples leave too much room for interpretation.

We need real-time control that shapes exactly how agents build on our framework.

The MCP Server

Our solution is a custom-built MCP (Model Context Protocol) server - a state machine that teaches agents how to build on our framework and PaaS.

Enforced Sequential Workflow

The MCP server implements a strict state machine: Login → PRD → Architecture Manifest → TDD Cycle → Deploy → Check → Ship. Agents can’t skip steps or jump ahead. This sequential enforcement removes decision paralysis and stops premature coding. Every application follows the same proven path.

Smart Failure Handling

When things go wrong, agents don’t flail around debugging. The state machine includes a failure loop that sends agents back to the TDD cycle. Failed deployments or validations automatically cause a return to tests, maintaining discipline even during debugging. This prevents the chaotic “fix-and-pray” debugging that agents often fall into.

Test-Driven Development Loop

The TDD cycle itself is a controlled loop: Tests → Code → Refactor → Iterate. Agents write tests first, implement code to pass those tests, then refactor - but can’t modify the tests themselves during refactoring. The iteration continues until all tests pass. No moving forward with broken tests, no building technical debt.

Validation Gates

Before any code ships, the MCP enforces endpoint testing and validation. The agent must prove the system works end-to-end, beyond unit tests. This final gate catches integration issues that might slip through component testing.

Structured Updates

Even modifications to existing systems follow the same disciplined path through an “Update” flow. This prevents cowboy coding on production systems - every change goes through the same quality controls.

The MCP has complete knowledge of our system components, making sure applications use built-in features rather than reinventing them. By controlling the development flow, we transform unpredictable agent behavior into reliable, repeatable software delivery.

We built the first framework and system designed specifically for AI agents to ship production applications. Built from the ground up for how agents actually work, rather than treating them as an afterthought.

The combination of opinionated design, versioned infrastructure, complete observability, intelligent annotations, pre-built building blocks, and guided development through MCP creates something new: a complete agent development system. Coding agents no longer waste time building basic infrastructure or wandering off into hallucinated solutions. They ship real applications, quickly and reliably.

These go beyond MVPs or demos. Our PaaS runs on a global edge network, meaning applications built by coding agents are production-ready from day one. They auto-scale to serve thousands of concurrent users without any additional configuration. The same application a coding agent builds in development deploys directly to handle real-world traffic at scale.

This framework represents the missing link in current agentic workflows. While everyone talks about AI agents changing software development, we built the infrastructure that makes it actually happen. Agents can now go from user requirements to deployed, scalable applications in hours, not weeks.

The age of AI-first development has arrived. We just needed to build the right framework for it.

Get Started in 5 Minutes

Ready to see your AI agent ship production code? The MCP server works with both Claude Code and Gemini CLI, guiding your agent through building real applications.

Step 1: Install the Raindrop CLI

npm install -g @liquidmetal-ai/raindrop

Step 2: Install the MCP for your agent

# For Claude Code
$ raindrop mcp install-claude

# For Gemini CLI
$ raindrop mcp install-gemini

Step 3: Start building Sign up at liquidmetal.ai to get your API credentials and watch your agent build its first production application today.

Subscribe to our newsletter

← Back to all Blogs & Case Studies