Search the Practicum

Type two letters · ↑↓ select · Enter open · Esc close

Type at least two letters.

Simple agents for personal and professional work

A living Practicum page. Checked: 2026-08-18. Agent features and access change every month, so check the date. The Book 1 route for individuals and specialists starts below. The Book 2 route for business platforms and rollout order is at the end of this page.

An agent is an AI system that does more than answer a question. It can work through several steps: search, fill in a form, send something, or combine results. The Book 1 route is simple. Set up one agent for a work task and one for your personal life. You are not trying to build an "agent company."

What you can use today

Tool What it is for
Gemini Spark by Google Personal agent for multi-step work and connected apps; access depends on country and plan
ChatGPT Work / cloud browser Multi-step research and browser actions; the former agent mode is no longer available as a separate mode
Zapier agents Workflows that connect several apps
n8n (self-hosted) Connect email, spreadsheets, and messaging apps into a workflow without writing code, on your own server
Hermes Agent (Nous Research) A self-hosted runtime that remembers your projects and writes its own reusable skills. MIT licensed, launched Feb 25, 2026. Needs a server and some technical judgment

A more advanced reference

Autonomous agent runtimes are a separate category. Hermes Agent and OpenClaw run on your own infrastructure, keep persistent memory, and learn through a loop in which a correction becomes a reusable skill. Their architecture connects to the harness from Chapter 6. See Autonomous agents: runtimes and frameworks.

Of the two, the author prefers Hermes as the easier starting point. See the author's stack. One security note is worth carrying with you: an always-on agent that reads messages from the outside world is a new attack surface, and 2026 produced published research on prompt injection that survives in an agent's own memory. Give an agent the narrowest access that still lets it finish the job.

Where to start as an individual

  • For work: choose one repetitive task that annoys you, such as sorting email or preparing a morning summary. Build one agent around it and take the workflow to a usable result. Check the result by asking, "Show me what this conclusion is based on."
  • For yourself: use AI for everyday life to explore agents for schedules, travel, and household tasks.

Discipline: an agent needs boundaries and review. Do not give it access to sensitive information until you understand what leaves your environment and where it goes. One workflow that you have tested is better than a plan to "automate my whole life."

Update rule: review every month and mark anything outdated. Version: 2026-08-18.

Agent platforms: where companies start (Book 2)

The individual route is above. This section explains what a team platform usually contains and how companies often roll it out. See Book 2, Chapter 3, "Teams with Agents," for the Agent Contract, the harness, and least privilege.

What an agent platform contains

A modern agent platform is a set of layers, not one tool. The three large clouds package those layers differently, which is the clearest way to see what a platform actually owes you.

Layer Purpose Where it sits on the big three
Agent builder Visual builder plus code Vertex AI Agent Builder; Azure AI Foundry; on AWS, the open Strands Agents SDK
Managed runtime Somewhere the agent actually runs, with sessions and identity Bedrock AgentCore; Azure AI Foundry Agent Service; Vertex AI Agent Engine
Tools and integrations A connector layer, now largely on the Model Context Protocol (MCP) MCP support on all three
Code execution A sandbox where the agent can run scripts Code interpreter tools on all three
Document search Retrieval augmented generation (RAG) over a company knowledge base Bedrock Knowledge Bases; Vertex AI Search; Azure AI Search
Web search Live web search Grounding tools, priced per query
Model choice Several large language models for different tasks Bedrock lists the widest catalog: more than 40 models from eight providers, checked Aug 2026
Streaming Low-latency voice and chat Realtime and live APIs on all three

Product names differ and are renamed often. The structure stays familiar: a builder, a runtime, tools, memory or search, models, and an execution environment. AWS is the clearest example of how layered this has become. It ships three separate products, the Kiro agentic development environment, the Bedrock AgentCore runtime, and the open-source Strands Agents SDK, where Microsoft and Google each lead with one front door.

Keep data inside and run computation outside when appropriate. Sensitive data stays in your own environment, either on premises or in a private cloud. The platform provides compute through an API. This is also a foundation for least privilege, which means giving each agent only the access it needs.

First, the honest scale of this

Before copying anyone's rollout, look at how few organizations have actually finished one. In McKinsey's State of AI survey, published Nov 5, 2025, 88% of organizations use AI regularly in at least one business function and 62% are at least experimenting with agents, but only 23% are scaling an agentic system in any function, and in any single function it is no more than 10%. Gartner expects more than 40% of agentic AI projects to be cancelled before 2027, citing unclear ROI, rising costs, and weak risk controls.

Read that as a working instruction rather than a warning. The gap between experimenting and scaling is where most of the money goes, and the sequence below is built to get you across it.

A typical rollout sequence

This pattern appears in retail, service businesses, and fintech, regardless of vendor.

Stage What the team does Time frame Risk
1. Content wins Summarize reviews, prepare drafts, and build FAQs 2 to 4 weeks Low
2. Company RAG Build an agent over the internal knowledge base 1 to 2 months Medium, because chunk quality matters
3. Search and recommendations Expand queries and personalize selections 2 to 4 months Medium
4. Complex machine learning Build recommendation systems and predictive analytics 6 months or more High without good data

Sequence rule: start with a task that can produce a result in two weeks and will not break anything if the agent makes a mistake. Move to complex machine learning after the team has learned to work with agents and can measure the outcome, not only the output.

An agent as a team member: the Claude Tag example (2026)

In June 2026, Anthropic placed Claude directly inside Slack. It was presented as a member of a work channel, not a personal assistant in a separate window. Each channel had one agent, shared memory, separate permissions, and logs that showed who assigned each task. The agent worked asynchronously.

Reports about the system said that Anthropic's product team used an internal version of this agent to write about 65% of its code. Treat that number as a vendor claim, not an independently verified metric.

The useful lesson is not "install the same product." Value moves toward the person who maintains the team's working context: its vocabulary, decisions, and priorities, which the agent builds up over time. This matches Book 2, Chapter 3. An agent is a participant with permissions, boundaries, and an owner under an Agent Contract. Shared memory and a shared operating environment can become a real competitive moat. Product names change. The principle does not.

Caution: check the source behind specific figures from vendor presentations. These numbers are often not publicly verified. The rollout sequence can be reproduced, but you should confirm any return on investment (ROI) figure in a public report.

Update this section when new cases appear with public sources.

Sources checked on 2026-08-18

Agent capabilities and rollout practices change quickly. Use these sources to check the latest details:

Use adoption percentages only when they come from reports such as McKinsey or Gartner. A vendor demo is a claim from that vendor, not an established fact.

Simple agents for personal and professional work