Modern Agentic AI Applications
A collection of agentic AI engineering work, kept deliberately honest about its status. In client work I have shipped MCP servers and clients and production agents on AWS Bedrock AgentCore; the items on this page are the experiments and prototypes behind that practice, not production deployments.
Problem statement
Problem statement
Agent demos are easy to build and hard to make dependable. The interesting engineering questions are about boundaries: which tools an agent may call, when a human must approve, and how to tell whether the system is getting better.
Business challenge
Business challenge
Agent frameworks change quickly and most public examples skip the hard parts — failure handling, evaluation, permissions and cost. Understanding those parts requires building small, focused systems and observing them.
Solution architecture
- 01
Each experiment isolates one question (for example: how should an MCP server expose a read-only data source?).
- 02
Experiments are kept small enough to reason about, with notes on what worked and what did not.
- 03
Patterns that hold up are carried into client work; the rest are documented as lessons.
01Interface
User / workflow trigger
Chat, API call or schedule
02Agent runtime
Planner
Decides the next step
Tool router
Validates and dispatches calls
03Capabilities
MCP servers
Data sources & actions
Retrieval
Knowledge base search
APIs
Internal & external services
04Control
Human approval
For consequential actions
Tracing & evaluation
Every step recorded
Workflow
- 1
Receive a goal
The agent gets a task and the set of tools it is permitted to use.
- 2
Plan
It decides the next action based on the goal and what it has learned so far.
- 3
Call a tool
The router validates the arguments and executes the call.
- 4
Observe
The result is added to the agent's working context.
- 5
Check in
Consequential actions pause for human approval.
- 6
Finish & trace
The final answer is returned and the full trace is stored for evaluation.
Technology stack
- Python
- Model Context Protocol (MCP)
- LangGraph
- LangChain
- AWS Bedrock AgentCore
- Langfuse
- Tool calling
- FastAPI
- Azure OpenAI
Implementation details
- Tool schemas are strict and validated before execution; a bad call is rejected with a useful error the model can recover from.
- Read and write tools are separated, and write tools require explicit approval.
- Every run is traced so behaviour can be compared between prompt and model versions.
Engineering challenges
Knowing when to stop
Agents loop. Step limits, budgets and clear completion criteria are essential.
Evaluating behaviour, not just answers
A correct final answer reached through a risky sequence of actions is still a failure.
Experiments and prototypes
MCP-powered AI applicationsPrototype
Building MCP servers and clients that expose data sources and actions to an LLM through a standard protocol, with attention to permissions and schema design.
- MCP
- Python
- FastAPI
Tool-calling systemsPrototype
Strictly-typed tool definitions, argument validation and error messages that let a model recover from a bad call instead of failing the run.
- Tool calling
- Pydantic
Multi-agent collaborationExperiment
Orchestrator-and-specialist patterns: how to split work, share context and merge results without duplicated or contradictory output.
- Agent orchestration
- LangChain
Human-in-the-loop systemsExperiment
Approval checkpoints for consequential actions, and how to present an agent's plan so a person can review it quickly.
- Workflow design
- FastAPI
AI workflow automationExperiment
Using an LLM for the judgement steps inside an otherwise deterministic workflow, rather than handing the whole process to an agent.
- Python
- Queues
Enterprise knowledge assistantsExperiment
Retrieval-backed assistants with source citations and access control on the retrieval layer.
- RAG
- Vector search
Autonomous research agentsPlanned
Long-running agents that gather, cross-check and summarize sources, with a focus on verification rather than volume.
- Agents
- Search tools
Business value and results
Business value
- Practical understanding of where agents help and where a simple pipeline is the better choice.
- Reusable patterns for permissions, approval and tracing.
Measured results
Experiments and prototypes only. None of these are claimed as production deployments.
Screenshots, code and demo
Screenshots
Will be added as the implementation progresses.
Source code
Repository not yet public.
Live demo
No public demo. Ask for a walkthrough.
Have an AI or Data Engineering challenge? Let's turn it into a working product.
Taking on a limited number of AI and data engineering projects. Typical reply within 2 business days.