Research documents, specifications, and reference implementations for building production-grade AI agents.
This repo deliberately follows a spec-first, implement-second approach. Each agent goes through a multi-stage pipeline before code is written:
- Research & domain analysis — understand the problem space, talk to domain experts, study existing workflows
- Specification — write a comprehensive spec covering architecture, data models, tool interfaces, edge cases, and compliance requirements. Specs are 1,000-3,800 lines each — detailed enough to be implementable by any developer.
- Reference patterns — distill agentic design patterns, failure modes, and SDK capabilities into reusable knowledge that applies across all agents
- Implementation MVP — build the working agent with MCP servers, tests, web UI, and interactive copilot
We currently have 10 specs and 2 implementations because:
- Writing a thorough spec surfaces 80% of the hard problems before a single line of code is written
- Specs are refined through multiple rounds — incorporating domain expertise, legal requirements, and real-world constraints
- Implementation only starts once the spec is stable and the architectural decisions are validated
- This prevents the most expensive kind of rework: building the wrong thing
The remaining 8 specs are queued for implementation based on customer priority and domain readiness.
├── reference/ # Agentic knowledge base — patterns, playbooks, failure modes
├── specs/ # Real-world agent specifications (10 agents)
├── implementation/ # Working code — TypeScript, MCP servers, web UI, tests (2 agents)
| Document | Description |
|---|---|
| Deep Agent Infrastructure Playbook | Comprehensive guide covering shared infrastructure patterns, architectural decisions, and implementation techniques for building deep agents. Covers tool design, MCP integration, orchestration, error handling, and deployment. |
| Agent Framework Implementation Playbook | Technical guide for implementing agents across 5 frameworks — Claude Agent SDK, LangChain/LangGraph, CrewAI, Mastra, and AWS Bedrock AgentCore. Includes framework selection matrix, reference implementations, hosting patterns (Docker/ECS/Fargate/Serverless), and migration guides. |
| Agentic Design Patterns Reference | Distilled patterns from Google's 482-page guide — prompt chaining, routing, parallelization, reflection, tool use, planning, multi-agent, memory, guardrails, reasoning. With priority implementation table. |
| Common Agent Failure Modes | 10 failure modes with symptoms, root causes, and mitigations — recursive loops, tool over-reliance, cascade failure, hallucinated results, context saturation, prompt injection, over-confident risk, state without rollback, blind retry, information hoarding. |
| Claude Agent SDK Patterns | Subagents (parallel execution, context isolation, tool restrictions), Slash Commands (built-in + custom with arguments/bash execution), and Agent Skills (autonomous invocation, SKILL.md format). With application examples for our real estate agent. |
| Agent | Description |
|---|---|
| Lab Report Intelligence Agent | Automates post-analyzer result validation and intelligent report generation for diagnostics labs — reference range checks, delta checks, clinical pattern detection, AI-powered interpretive comments, and tiered autonomy (auto-validate normals, escalate abnormals). |
| MR Copilot Agent | AI assistant for Medical Representatives in pharma — pre-call intelligence (doctor profiling, RCPA trends, visit history), in-call support (product Q&A, objection handling, clinical evidence), voice-to-DCR automation, route optimization, UCPMP compliance tracking, and manager coaching signals. |
| ABHA Health Record Agent | Personal local agent that manages your ABHA (Ayushman Bharat Health Account) — browser automation for portal navigation, health record download and organization, health summaries, consent management, and full audit trail with screenshots. |
| Agent | Description |
|---|---|
| SOC 2 Compliance Agent | Manages the complete SOC 2 compliance lifecycle — scoping, control mapping, policy generation, evidence collection, gap assessment, remediation tracking, auditor coordination, and continuous monitoring. |
| Statutory Compliance Calendar Agent (India) | Tracks, alerts, prepares, and assists in filing 30+ recurring statutory compliance obligations across GST, Income Tax, TDS/TCS, PF, ESI, Professional Tax, ROC/MCA, and state-level filings for Indian businesses. |
| Legal/Contract Intelligence Agent (India) | Indian contract review agent — reads contracts holistically, flags risks against Indian law (Contract Act, DPDPA, Stamp Act, Labor Codes, FEMA), explains in plain language for business teams, generates redlines and negotiation playbooks, tracks versions through negotiation rounds, and calculates state-wise stamp duty. Legal KB in PostgreSQL + pgvector. |
| Agent | Description |
|---|---|
| Vendor Onboarding Agent (India) | Handles end-to-end vendor onboarding for Indian businesses — document collection, government API validation (GSTIN, PAN, Udyam, MCA, Bank), compliance screening, risk scoring, approval workflows, and ERP master creation. |
| Cloud Cost Optimization Agent (AWS) | Investigates cost spikes, detects and cleans up waste, rightsizes instances with auto-rollback, optimizes Savings Plans, schedules dev/staging environments, and analyzes data transfer costs — across AWS Organizations with tiered autonomy. The agent that executes, not just recommends. |
| Agent | Description |
|---|---|
| Real Estate Transaction Agent (Gujarat) | End-to-end property purchase companion for Gujarat — browser automation for 6 government portals (AnyRoR, Gujarat RERA, eCourts, GARVI, SMC, GSTN), due diligence with dispute search, builder agreement review, total cost breakdown (jantri vs market rate), stamp duty calculation, registration guide, and immutable purchase dossier with timestamped evidence. |
| Agent | Description |
|---|---|
| Personal Accounting Agent | Research document for an AI-powered personal accountant that handles financial data processing, categorization, reconciliation, reporting, and insights. |
Working code with MCP servers, tests, web UI, and interactive copilot.
| Agent | Stack | Tools | Features |
|---|---|---|---|
| Real Estate Transaction Agent | TypeScript, Claude Agent SDK, Hono, Vite+React | 4 MCP servers (24 in-process tools + Playwright browser) | Web UI (wizard → chat), 11 slash commands, 5 subagents, 5 skills, reflection/critic agent, Playwright MCP fallback for CAPTCHA |
| Legal Contract Intelligence Agent | TypeScript, Claude Agent SDK | 3 MCP servers (14 tools) | CLI copilot, 11 slash commands, 4 subagents, 5 skills, reflection/critic agent, multi-state stamp duty |
Contributions are welcome! Whether it's a new agent spec, improvements to existing ones, or additions to the reference material.
- Fork the repository
- Create a branch for your contribution (
git checkout -b feature/your-agent-spec) - Follow the existing structure:
- Agent specs go in
specs/— use existing specs as a template for format and depth - Reference/knowledge docs go in
reference/ - Working implementations go in
implementation/
- Agent specs go in
- Submit a pull request with a clear description of what you're adding
- Agent specs that solve real business problems with enough depth to be implementable
- Reference material that distills practical, non-obvious insights from building agents
- Implementation code with tests and clear MCP tool interfaces
- Bug fixes or improvements to existing specs based on real-world testing
This project is licensed under the MIT License — free to use, modify, and distribute.
- Dhaval Nagar — dhaval@appgambit.com
- Antigravity Apps — Product studio building AI-native applications
- AppGambit — Cloud consulting and engineering