AI agents are systems that pursue goals on their own — they plan, use tools, and take actions, rather than simply answering a question. If the chatbot era put intelligence behind a text box, the agent era puts it behind a to-do list. You stop prompting turn by turn and start delegating outcomes.
That shift — from answering to acting — is the most important thing happening in software right now, and it's why capital and talent are pouring into the layer above the foundation models. We track it daily across the AI Agents theme, the podcasts where these founders explain themselves, and the papers pipeline.
Agent vs. Chatbot vs. Copilot
| Chatbot | Copilot | Agent | |
|---|---|---|---|
| You provide | A question | A task, in context | A goal |
| It returns | An answer | A suggestion you approve | A completed outcome |
| Autonomy | None | Low — you stay in the loop | High — acts on its own |
| Can work unattended? | No | Barely | Yes (that's the point) |
| Example | ChatGPT Q&A | GitHub Copilot inline | Manus, Claude Code, Devin |
The line that matters is the last row: can it do useful work while you're not watching? That single property is what separates a genuinely agentic product from a chatbot wearing the word "agent" in its marketing.
The Four-Layer Agent Stack
1. The model — the reasoning core. Every agent runs on a foundation model (GPT, Claude, Gemini, open models) doing the planning and decision-making. Notably, most agent companies don't train their own — they orchestrate someone else's. That's why a small team could ship Manus without a billion-dollar training run. How a model like ChatGPT works is the foundation under this whole stack.
2. Tools & actions. What turns a model into an agent: a browser, code execution, file access, APIs, and the ability to call them. "Tool use" and "function calling" are the plumbing; the open standard most teams now build against is the Model Context Protocol (MCP).
3. Memory & orchestration. The loop that keeps an agent coherent across many steps — short-term scratchpads, long-term memory, retries, and the planner that decides what to do next. This is the hardest and least-solved layer, and where most agents quietly break.
4. The interface & guardrails. How you delegate and supervise — approvals for consequential actions, sandboxing, scoped permissions. The companies that win enterprise will win here, because trust, not raw capability, is the bottleneck.
Why Agents Took Off in 2025–2026
Three things converged. Models got good enough at reasoning and tool use to chain steps reliably-ish. The standards matured (function calling, MCP) so agents could plug into real systems. And the economic logic clicked: if a chatbot is worth $20/month, an agent that does the work itself is priced against labor — a vastly larger market. That's why the agent layer attracted the funding it did, and why our AI Agents theme now spans 175+ tracked companies, from general agents to vertical ones in browser automation and app development.
The Honest Counterweight
The demos are intoxicating; the deployments are harder. The core problem is compounding error: chain ten steps that are each 90% reliable and the whole task succeeds only ~35% of the time. Add prompt-injection risk, the cost of autonomous wrong actions, and uneven tool reliability, and you get the real 2026 status: agents are production-ready for narrow, supervised, well-instrumented tasks — and oversold for everything else. The interesting signal isn't who demos an agent; it's who deploys one that survives contact with real users. That's what the live map below is for.
The Live Map
AI Agent Companies by Signal Volume
Live membership of the ai-agents, agentic-browser-automation & ai-agentic-app-development themes · ranked by extracted signals
- 01Anthropiclast seen JUN 13441 signals
- 02OpenAIlast seen JUN 13379 signals
- 03Googlelast seen JUN 11138 signals
- 04Metalast seen JUN 13112 signals
- 05Salesforcelast seen JUN 1142 signals
- 06Manuslast seen JUN 1123 signals
- 07Cognitionlast seen JUN 1019 signals
- 08Legoralast seen JUN 917 signals
- 09Sierralast seen JUN 1114 signals
- 10Generalist AIlast seen JUN 1110 signals
- 11Flourishlast seen JUN 97 signals
- 12LangChainlast seen JUN 87 signals
- 13Thinking Machines Lablast seen JUN 67 signals
- 14Townlast seen JUN 106 signals
- 15Factorylast seen JUN 65 signals
- 16Reactorlast seen JUN 15 signals
- 17Beaconlast seen JUN 104 signals
- 18Periodic Labslast seen JUN 104 signals
- 19Gradient Labslast seen JUN 84 signals
- 20Sarislast seen JUN 44 signals
Going Deeper
- The breakout agent: What is Manus? — the viral general agent, with our live signal feed on it.
- The model layer underneath: How does ChatGPT work? — the reasoning core agents orchestrate.
- The adjacent frontier: What is physical AI? — agents that act in the physical world, not just the browser.
- The alternatives & competitors: ChatGPT alternatives · OpenAI competitors · Anthropic competitors.
- The research: our papers pipeline summarizes the top agent and reasoning papers daily.
The agent field moves weekly — who shipped, who raised, who quietly stalled. The free Teahose daily digest tracks every agent launch, funding round, and deal across 40+ podcasts, 20+ newsletters, and the day's research, distilled into one morning email. Subscribe free and watch the category move in real time.
Definitions are stable; the live company map is as of June 14, 2026, and updates continuously.
Frequently Asked Questions
What are AI agents?
AI agents are software systems that pursue a goal autonomously. Instead of returning a single answer like a chatbot, an agent takes an objective ("book this trip," "triage these tickets," "research these companies"), breaks it into steps, uses tools — a web browser, code execution, APIs, a file system — and loops through plan-act-observe-correct until the task is done. The defining trait is agency: it decides what to do next, not just what to say next.
What is the difference between an AI agent and a chatbot?
A chatbot is reactive and conversational — you ask, it answers, you steer every turn. An agent is goal-directed and autonomous — you delegate an outcome and it takes multiple actions on its own to reach it. A useful test: if the system can do something while you're not watching (browse 30 sites, run code, send requests, then hand you a finished deliverable), it's acting as an agent. If it only produces text in response to each prompt, it's a chatbot, even a very smart one.
What are the best examples of AI agents?
Consumer-facing: Manus (general task agent), OpenAI's and Anthropic's computer-use and "operator"-style agents, and browser agents that navigate the web for you. Developer-facing: coding agents like Claude Code, Cursor's agent mode, and Devin-style autonomous engineers. Enterprise: customer-support agents (Sierra), workflow and back-office agents, and agentic search. Teahose tracks 175+ companies in the AI Agents theme alone — the live map below ranks them by current signal volume.
Do AI agents actually work yet?
On narrow, well-scoped tasks with good tools — coding, structured research, repetitive web workflows — yes, increasingly well. On long, ambiguous, high-stakes tasks, not reliably. The honest failure mode is compounding error: a 90%-reliable step run ten times in a chain succeeds end-to-end only about a third of the time. That gap between an impressive demo and a dependable deployment is the central engineering problem of the category in 2026, and it's exactly what to watch when a company claims its agent is production-ready.
Are AI agents safe?
They introduce genuinely new risk. An agent with a browser, your credentials, and the ability to act can be steered by malicious content it reads on the web (prompt injection), can take costly wrong actions autonomously, and expands the data it touches. The mitigations — sandboxing, human-in-the-loop approval for consequential actions, scoped permissions — are improving but not solved. Treat an agent like a fast, capable, occasionally-overconfident new hire: useful with supervision, not yet trustworthy unattended.