AGIWORK / FIELD GUIDE

Build the ecosystem.
Keep it purposeful.

A practical guide to the layers, choices and tools behind an enterprise AI platform — from a first experiment to managed operations.

Explore the architecture Browse the toolkit

MY APPROACH

Start small. Add capability only when a real requirement justifies it.

THE IDEA / FROM MY REFERENCE ARCHITECTURE

One ecosystem.
Different starting points.

I use this map to separate the jobs an AI platform has to do: serving people, connecting models and knowledge, controlling access, and operating safely. The toolkit below is an adapted selection from my enterprise AI ecosystem catalogue.

You do not need every component. Begin with a defined use case and add layers as data sensitivity, users and operational requirements grow. The stages are a learning path, not a maturity certification or a prescription.

01 / ARCHITECTURE MAP

See how the pieces
fit together.

This is a functional map, not a required deployment topology. Each area links to examples and decisions in the toolkit below.

PEOPLE & ACCESS

Channels & interfacesChat, web, mobile, learning and business workflows Identity & accessSSO, roles, policies and secrets

APPLICATIONS

Applications & agentsAgent frameworks, workflows, human review and artifact registries Coding & development agentsDeveloper agents, coding environments and execution harnesses Safety controlsGuardrails, sensitive-data protection and tool permissions

CAPABILITIES

Knowledge & dataSources, processing, OCR, retrieval, vector and graph search Models & gatewayLocal and external models, serving and routing AI servicesText, audio, images and model development

CONNECTIONS

Integration & executionAPIs, MCP, agent communication, event bus and isolated execution Operations & securityEvaluation, tracing, monitoring, audit and response

FOUNDATIONS

Governance & FinOpsOwners, catalogues, policies, risk and cost visibility Enterprise systemsFiles, databases, documents, communications, CRM, Git and APIs

Identity, permissions, evaluation and oversight span the entire system. Components within a category are often alternatives, not a shopping list.

02 / A LEARNING PATH

From first experiment
to reliable delivery.

I group tools by the work you are doing. Some tools serve several stages; a later stage is not automatically better.

Start

Explore models, a simple interface and a small task.

Explore starting points ↗02

Build

Add agent logic, documents, retrieval or media only as needed.

Agents & knowledge ↗03

Integrate

Connect services and models with explicit access and controls.

Connections & safety ↗04

Operate

Evaluate, monitor and govern the system as usage grows.

Operations & governance ↗

03 / TOOLKIT

Choose by job,
not by popularity.

Official project links are starting points for evaluation. These projects overlap and evolve; check current releases, compatibility, maintenance, security and licence terms before adoption.

THE COMPLETE COLLECTION

Explore the wider agent ecosystem.

Browse my full directory by architecture role, from personal assistants and coding agents to models, knowledge and operations.

Explore all projects ↗

01 / START

Interfaces & first experiments

Give people a useful way to interact with a model and learn from a small, bounded use case.

MY TAKEStart with one interface and one model. A self-hosted interface does not automatically make an external model or its data processing private.

AI & agent interface

Open WebUI

A self-hosted interface for models, tools, knowledge and configurable agents.

Source-available; review licence and branding terms.Agent & workflow automation

n8n

Visually build workflows and AI agents that connect to services.

Fair-code; review the Sustainable Use License.Local model runtime

Ollama

Run supported models locally for experiments and development.

Model licences are separate.AI chat interface

LibreChat

A multi-user chat interface and integration starting point.

Check edition and deployed version.Prototype UI

Gradio

Build a simple web interface for a model or Python function.

Check dependencies and hosting.Web UI library

React

Build a custom web interface for AI applications.

You provide the application and its security controls.Shared learning environment

JupyterHub

Provide multi-user notebook environments for learning and exploration.

Configure user isolation and resource limits.Learning platform

Moodle

Organise courses, learning content and activities.

Review extensions and operational requirements.Visual workflows

Node-RED

Connect events, APIs and services in low-code flows.

Review third-party nodes.

02 / BUILD

Agents & orchestration

Build a single agent or workflow with an explicit task, tools, state and clear boundaries.

MY TAKEA model runtime, an agent framework and a workflow engine have different jobs. Choose the simplest fit; multi-agent coordination is not a prerequisite.

AI & agent interface

Open WebUI

Configure assistants and connect models, tools and knowledge through a self-hosted interface.

Source-available; review licence and branding terms.Agent & workflow automation

n8n

Create AI agents and automated workflows with visual building blocks.

Fair-code; review the Sustainable Use License.Personal agent platform

OpenClaw

Run a personal assistant that connects models, tools and messaging channels.

Review workspace permissions and access to connected accounts.Personal agent platform

Hermes Agent

Run a tool-using assistant with persistent context, skills and messaging integrations.

Review skills, tool access and the selected model provider.Personal agent framework

nanobot

Build or run a lightweight personal agent with tools, memory and chat integrations.

Review exposed tools, credentials and deployment settings.Agent library

smolagents

A lightweight Python starting point for tool-using agents.

Code-executing agents require isolation.Visual agent builder

Flowise

Compose AI and agent workflows through a visual interface.

Review community and commercial editions separately.Agent framework

Pydantic AI

Python agents with typed inputs, outputs and validation.

Check model and provider terms.Stateful orchestration

LangGraph

Build stateful, controllable agent workflows.

CrewAI

Coordinate agents and tasks where roles and handoffs are useful.

Review core and hosted offerings separately.Agent toolkit

Google ADK

A code-first toolkit for building and evaluating agents.

Cloud integrations may have separate terms.

02 / BUILD

Coding agents & development harnesses

Give development agents a bounded workspace to inspect code, propose changes, run tools and support delivery.

MY TAKEA coding agent combines a model with an execution harness and tools. Compare workspace permissions, model access, review workflows and data handling before adoption.

Coding agent

OpenCode

Use an open coding agent to work with a codebase from developer interfaces.

Check provider terms and execution permissions.Coding agent

Claude Code

Work with an agent that reads, changes and tests code through a terminal and other integrations.

Review Anthropic service and workspace permissions.Coding agent & harness

OpenAI Codex

Use OpenAI’s coding agent across app, CLI, editor and programmatic workflows.

Open-source harness and hosted services have separate terms.Coding agent toolkit

Pi

Use an extensible terminal coding agent or its agent toolkit as a building block.

Inspect extensions and configure appropriate execution boundaries.Agentic development environment

ZCode

Use ZCode’s agent to work with files, commands, models and development tasks.

Review product, provider and data-handling terms.Agentic development platform

Google Antigravity

Use agents in development workflows across an editor, terminal and browser.

Review access, availability and product terms.Agent harness

DeepSeek Harness

Assemble plugin-based agent workflows with sessions, tools and a web interface.

Review plugins, model access and permissions before use.Terminal coding assistant

Aider

Pair with an AI assistant to edit code in a local Git repository.

Review model-provider terms and changes before merging.

02 / BUILD

Documents, retrieval & knowledge

Turn approved source material into searchable evidence that an application can cite.

MY TAKEFor RAG, check extraction quality, access permissions, retrieval relevance and source freshness. A vector database is not an access-control system by itself.

Document parsing

Docling

Convert supported documents into structured content; OCR options are available.

Check OCR engines and model licences.Data integration

Airbyte

Move supported data from sources into destinations for processing.

Check connector and edition terms.OCR

Tesseract

Recognise text in images and scanned pages.

Quality depends on source and language data.Vector search

pgvector

Add vector similarity search to PostgreSQL.

A pragmatic option when PostgreSQL is already in use.Vector search

Qdrant

A dedicated vector search engine.

Compare filtering, operations and scale requirements.Vector database

Milvus

Vector search infrastructure for larger deployments.

Distributed deployment adds operational work.Search engine

OpenSearch

Explore full-text and vector search for retrieval.

Check version, plugins and operational requirements.Graph database

Neo4j Community

Model and query relationships as a graph.

Community and enterprise editions have different terms.RAG framework

Haystack

Compose retrieval and AI pipelines from components.

Evaluate retrieval and answers separately.

02 / BUILD

Audio, images & model development

Add media processing or adapt a model when the problem actually requires it.

MY TAKEStart with existing models and a test set. Fine-tuning, visual workflows and transcription introduce additional compute, data and model-licence decisions.

Speech recognition

Whisper

Transcribe speech with supported open model weights.

Review model and data terms.Audio & video processing

FFmpeg

Process and convert supported audio and video formats.

Enabled libraries and build options affect licence obligations.Visual media workflows

ComfyUI

Compose node-based image and other media workflows.

Extensions and models carry their own terms.Model training

Hugging Face TRL

Train or fine-tune language models with supported methods.

Requires data, evaluation and compute planning.Synthetic data

SDV

Explore synthetic tabular data generation.

Synthetic data is not automatically private; evaluate disclosure risk.

03 / INTEGRATE

Model gateways & serving

Route applications to approved local or external models, and select serving infrastructure appropriate to demand.

MY TAKEOllama is a useful local starting point. vLLM and KServe address different serving and deployment needs; benchmark on your own hardware and models.

Local model runtime & API

Ollama

Run and serve supported models locally for applications and development.

Review each model’s licence and hardware needs.Model gateway

LiteLLM

A proxy and unified interface for supported model providers.

Review version, edition and integrations.Inference serving

vLLM

Serve supported models with an inference engine and API.

Check hardware and model compatibility.Inference serving

SGLang

A runtime for serving supported generative models.

Benchmark with your workload.Model serving on Kubernetes

KServe

Serve and manage model workloads on Kubernetes.

Requires Kubernetes operations.

03 / INTEGRATE

Tools, APIs & workflows

Connect approved systems through controlled interfaces, and make long-running actions observable and recoverable.

MY TAKEMCP is a protocol, not a security boundary. Use least-privilege credentials, approval for consequential actions and isolation for untrusted execution.

Tool integration protocol

Model Context Protocol

A specification for connecting applications to tools and context.

Review each server, permissions and transport.Skills specification

Agent Skills

A portable format for packaging agent instructions and supporting resources.

Treat external skills as untrusted until reviewed.Durable workflows

Temporal

Coordinate long-running workflows and recovery.

Design idempotency and human review.Messaging

NATS

Messaging infrastructure for services and event-driven systems.

Protect identities and network access.Isolation

Firecracker

MicroVM technology for isolated workloads.

Isolation requires secure configuration and operations.

03 / INTEGRATE

Identity, policy & guardrails

Protect access to the platform, models, data and tools using several complementary controls.

MY TAKEIdentity, secrets, authorization, data protection and AI guardrails are separate layers. No prompt filter alone prevents every injection or data leak.

Identity & SSO

Keycloak

Manage identity and access for applications.

Integrate with existing identity policies.Secrets management

OpenBao

Manage secrets and encryption with identity-based access.

Plan deployment and recovery.Policy engine

Open Policy Agent

Evaluate policies separately from application code.

Define and test enforcement points.Sensitive-data protection

Microsoft Presidio

Detect and help anonymise sensitive information.

Detection is imperfect; test on your data.AI guardrails

NeMo Guardrails

Add programmable controls around LLM applications.

Combine with application security and evaluation.

04 / OPERATE

Evaluation & AI security

Test both useful behaviour and failure modes before expanding access or autonomy.

MY TAKECreate a representative test set and inspect retrieval, answers, permissions, unsafe tool calls and regressions. Human review remains important for consequential actions.

Evaluation & red teaming

promptfoo

Evaluate prompts and test LLM applications for security risks.

Review configuration and data sent to providers.LLM vulnerability probing

garak

Probe language models for a range of failure modes.

A scan is not a guarantee of safety.Supply-chain scanning

Trivy

Scan supported software artifacts and dependencies.

Use alongside patching and review.

04 / OPERATE

Tracing, metrics & operations

Make model calls, tool use, latency, errors and operational costs visible to the teams responsible.

MY TAKETracing may contain sensitive inputs or outputs. Minimise logged data, control access and define retention before using production data.

LLMOps & model lifecycle

MLflow

Track experiments, models and supported AI traces.

Check deployment and access controls.LLM observability

Langfuse

Inspect traces and evaluate LLM applications.

Review self-hosted and commercial features.Telemetry

OpenTelemetry

Collect and process supported traces, metrics and logs.

Avoid logging unnecessary sensitive data.Metrics

Prometheus

Collect and query operational metrics.

Metrics complement traces and logs.Dashboards

Grafana

Visualise and investigate operational data.

Review the licence and deployment edition.

04 / OPERATE

Governance, catalogues & costs

Assign owners and policies to data, models and services; make adoption and cost decisions reviewable.

MY TAKEA data catalogue, a policy engine and a cost monitor support governance but do not create compliance by themselves. Controls depend on the organisation and jurisdiction.

Data catalogue

OpenMetadata

Catalogue data assets, metadata and lineage.

Define owners and classification processes.Cost visibility

OpenCost

Measure Kubernetes workload costs and allocation.

Not a complete model-token billing system.Artifact registry

Harbor

Store and manage container images and other supported artifacts.

Review provenance and access policies.Artifact signing

Sigstore Cosign

Sign and verify supported software artifacts.

Use with a defined trust policy.

LICENSING / IMPORTANT DISTINCTION

Every project has its own terms.

I track the licences and product terms for the projects in this architecture and my agent directory. A repository licence does not automatically cover hosted services, dependencies, models, weights or datasets. Check the exact version and use case.

Project licence sources checked 2026-10-29. Standard labels reflect the reviewed root licence and GitHub's identification; exceptions and uncertainties are stated explicitly.

Some important distinctions

open-webui

Open WebUI License

Branding rules apply from v0.6.6; earlier code and small deployments have different terms.

Build the ecosystem.Keep it purposeful.

One ecosystem.Different starting points.

See how the piecesfit together.

From first experimentto reliable delivery.

Start

Build

Integrate

Operate

Choose by job,not by popularity.

Explore the wider agent ecosystem.

Interfaces & first experiments

Open WebUI

n8n

Ollama

LibreChat

Gradio

React

JupyterHub

Moodle

Node-RED

Agents & orchestration

Open WebUI

n8n

OpenClaw

Hermes Agent

nanobot

smolagents

Flowise

Pydantic AI

LangGraph

CrewAI

Google ADK

Coding agents & development harnesses

OpenCode

Claude Code

OpenAI Codex

Pi

ZCode

Google Antigravity

DeepSeek Harness

Aider

Documents, retrieval & knowledge

Docling

Airbyte

Tesseract

pgvector

Qdrant

Milvus

OpenSearch

Neo4j Community

Haystack

Audio, images & model development

Whisper

FFmpeg

ComfyUI

Hugging Face TRL

SDV

Model gateways & serving

Ollama

LiteLLM

vLLM

SGLang

KServe

Tools, APIs & workflows

Model Context Protocol

Agent Skills

Temporal

NATS

Firecracker

Identity, policy & guardrails

Keycloak

OpenBao

Open Policy Agent

Microsoft Presidio

NeMo Guardrails

Evaluation & AI security

promptfoo

garak

Trivy

Tracing, metrics & operations

Build the ecosystem.
Keep it purposeful.

One ecosystem.
Different starting points.

See how the pieces
fit together.

From first experiment
to reliable delivery.

Choose by job,
not by popularity.