An AI Tools Knowledge Map — Lining Up the Materials for Process-Improvement Consulting#

2026-06-14

Introduction#

For a while now, the author has felt a vague thirst. New AI tools pour out every week, yet I have not been able to lay out, on a single page in my head, “what tools exist, what each one is, and when and in what combination they are best used.” I use individual tools, but I had never placed them on a map and looked at them together.

This thirst has a purpose. The author wants to reach a level where I can propose and consult on improving work processes with AI. The first thing required for that is not a dazzling strategy, but lining up the materials. Only once I have a firm grasp of what tools exist and what kind of tool each one is can I build combinations and strategy on top of them.

So this piece is not a finished argument but a knowledge map. It is a living list that the author will review often to refresh memory, updating it whenever I try a tool firsthand or learn about a new one. Each tool carries its relative strength (what it does better than the alternatives) and the author’s current learning status (in use / have experience / no experience).

One caveat up front: the “relative strength” in the tables is not a claim verified by benchmarks, but a summary of vendors’ official positioning and industry consensus as of June 2026. The map of strengths shifts with every model release, so please read them as a starting point rather than a conclusion.

1. General-Purpose Chatbots / Assistants#

General-purpose chatbots handle almost everything conversationally in one place — Q&A, drafting, summarization, translation, even code assistance — making them the most basic tool group and everyone’s entry point. They are also the first step when discussing AI-driven work improvement.

There is one important trend here. By 2026, the fundamentals of the top models (reasoning / multimodality / long context) have rapidly leveled up, and the big three (ChatGPT / Gemini / Claude) all now support context windows in the million-token range. As a result, declaring an absolute “A is simply better than B” has become increasingly difficult. So the basis for choosing has shifted from raw performance to relative strengths and ecosystem fit.

Reflecting this, the table below concisely captures what each tool is “relatively better at” rather than ranking them absolutely. Keep in mind these are positioning and consensus as of June 2026.

2. AI Coding Tools#

The core tool group for improving product development processes. Even within “coding AI,” the character varies quite a bit, and two axes make it easy to sort out. One is who drives — a spectrum from assistive tools (autocomplete) where the human leads and the AI helps, to autonomous agents that take on entire tasks. The other is where it runs — in your local editor·terminal, or asynchronously in the cloud. By these two axes, the tools split into the three branches below.

2-1. IDE / Editor Type#

A form where the human drives inside the editor, using the AI’s autocomplete, refactoring, and agent features together.

2-2. CLI / Terminal Agent Type#

A form strong at local·headless execution, handling an entire repository and automating work from the terminal.

2-3. Cloud / Autonomous Coding (App Generation)#

A form where the AI takes on work asynchronously given only a goal, or instantly produces an app·UI from a prompt alone.

3. Autonomous Agents#

A tool group that, rather than answering one-off queries, takes a goal, plans on its own, and executes multiple steps. If the coding agents above are specialized for the narrow domain of “development,” these aim to autonomously carry out broader general-purpose work — search, file manipulation, calling external tools, and more. From a consulting view, this is the area that gauges “can repetitive work be run without human intervention?” Because tools the author actually operates at work belong here, they are covered in depth later (Section 6).

4. Research / Knowledge Management#

Where general-purpose chatbots answer with “the world’s knowledge,” this tool group specializes in answering with evidence found within material you specify. By reducing hallucination (inventing facts that do not exist) and keeping sources traceable, it is strong for learning, research, and organizing the basis for decisions.

5. Image / Video / Audio Generation#

A tool group that weighs heavily in marketing and content-production processes. It can quickly mock up deliverables that used to be labor-intensive — ad assets, promo videos, narration — so the work-improvement payoff is highly visible here. Since the leader differs by domain, the domain is noted alongside.

6. Self-Hosted Autonomous Agents — Hermes & OpenClaw#

The author has hands-on experience with these two at work, so they can serve as living cases in consulting. They are therefore covered in greater depth.

Hermes (Nous Research)#

Hermes is an open-source autonomous agent released by Nous Research in February 2026. Its key differentiator is that it is neither a coding copilot nor a thin API-wrapper chatbot, but an always-on agent that lives on a server and gets smarter the longer it runs.

• A self-improving skills loop and persistent memory — it creates skills from experience on its own and accumulates an understanding of the user/project across sessions.
• A multi-platform gateway — it connects 20+ channels such as Telegram / Slack / Discord / WhatsApp / Email / CLI through a single gateway, and runs unattended automations (daily reports, etc.) via a cron scheduler.
• Self-hosting and data sovereignty — it can be deployed anywhere from a small VPS to a GPU cluster to serverless, and data never leaves your infrastructure. The model can be freely chosen from OpenRouter (200+ models) / OpenAI / your own endpoint, so there is no vendor lock-in.

The author uses Hermes at work for both product development and ad operations, and we are discussing running it on a shared server for team-wide use. Its single-gateway + multi-channel + cron structure suits a shared agent well, but going shared makes the key challenges clear: preventing dashboard exposure (it binds to localhost by default), running non-root with isolated sandboxes, governing the multi-channel and API keys, and designing permissions around “whose data / context is visible to whom.”

OpenClaw (OpenClaw Foundation)#

OpenClaw is an open-source (MIT) local-first, multi-channel agent. It is also the tool the author tried before adopting Hermes.

One common misconception needs correcting here. The claim that “OpenAI acquired OpenClaw” is inaccurate. In February 2026, OpenAI hired (an acqui-hire) OpenClaw’s sole developer to lead personal-agent development — it did not buy the company or the codebase. The OpenClaw software remains open source under the MIT license and has moved to an independent foundation (OpenClaw Foundation); OpenAI is merely a financial sponsor with no governance rights.

This distinction matters for assessing OpenClaw’s potential. Its potential may grow, but not because it was “absorbed into an OpenAI product.” Rather, it is because (1) the creator now has access to frontier models and compute, strengthening the project’s capabilities, (2) its architecture is signaled to influence the ChatGPT product, and (3) thanks to the foundation structure it remains open source, so enterprises can adopt it without fear of lock-in.

7. AI Delivery Environments — Managed AI Platforms#

An essential area when proposing AI environments to security-sensitive enterprises. Environments like AWS Bedrock, which serve frontier models securely, belong here.

Terminology First#

The phrase “AI delivery environment” gets the meaning across but is not standard terminology. The common terms are Managed AI Platform, Enterprise AI Platform, Foundation-Model Service (FMaaS), and model serving / inference platform. If picking a single phrase, “managed AI platform” or “foundation-model platform” is safe.

Correcting the Understanding of Bedrock#

It is easy to understand why security-sensitive enterprises prefer Bedrock-type services as “a dedicated line to the model vendor,” but the precise core is the following.

• Data does not leave the cloud boundary — because the model is hosted in the cloud, data is not transmitted directly to the model vendor.
• No training on your data / data residency — inputs and outputs are not used to train the base model and do not leave the designated region.
• Network isolation — VPC and private endpoints bypass the public internet. (The part closest to the “dedicated line” feeling.)
• The feature where “dedicated” is precisely right is Provisioned Throughput — an option to reserve specific model capacity for your exclusive use; this fits the dedicated-throughput concept better than networking does.

Major Platforms and the Deployment Spectrum#

Deployment forms make up the following spectrum in order of security strength.

Public SaaS API → managed cloud (Bedrock / Vertex / Azure) → private VPC → on-premise / air-gapped (Moving right, security rises, and so do cost and operational burden.)

Security Proposal Checklist#

No training on your data·zero retention, data residency, network isolation (VPC / private endpoints), access control (IAM / RBAC), encryption (at rest·in transit), audit logging, guardrails (content filtering·PII masking·prompt-injection defense), compliance certifications (SOC 2 / ISO 27001 / HIPAA / GDPR / public-sector FedRAMP·Korea’s CSAP), pricing model (on-demand vs Provisioned Throughput), and usage approach (RAG vs fine-tuning). Working through these lets you propose a trustworthy environment to security-sensitive enterprises.

8. One Layer Down — Developer Infrastructure#

If the tools above are finished “products,” this group is the foundation one layer down that connects, runs, and lets you assemble them directly. When consulting forces you to propose self-hosting or customization for security·cost reasons, the answer comes from exactly this layer.

9. The Learning Plan from Here#

With this map drawn, the next steps became clear.

First, fill in the blanks. Start by trying firsthand the tools that are currently “no experience” but high in recognition and real-world usage. Among chatbots, Perplexity; in coding, Windsurf·Devin; in media, Midjourney·ElevenLabs·Sora; and among platforms, hands-on practice with AWS Bedrock come first.

When studying each tool, organize it with the same template — what it is / what form it takes / strengths and limits / suitable work scenarios / security·cost characteristics / alternatives / a one-line impression from using it.

And the final destination of this map is the scenario × tool combination. Place work types (research / code development / content creation / ad operations / docs·reporting) on one axis and the client’s security tier (general startup / regulated industry / air-gap-requiring enterprise) on the other, and build a matrix where you can say “for this situation, this combination.” Only then will I be able to put the word “consulting” on my lips.

Closing#

The author believes a good proposal comes not from a dazzling conclusion but from well-organized materials. This piece is the first page of those materials lined up. If each tool is a dot on the map, the work ahead is to touch the dots firsthand and connect them with lines. And those lines, gathered together, will become the answer to “whom should I recommend which environment to.”

The author will keep this map open and update it often. The next piece will tackle the work scenario × security tier × tool combination matrix foreshadowed here.