This page is a landscape snapshot, not a live ticker. Product names and GA dates move. When in doubt, believe the vendor’s current docs over a blog post — including this one. Last editorial pass: August 2026.
The useful question is not “which logo is an agent?” It is: can this system take bounded actions with a trace, and can I stop it?
What “agent” means here
On Bye Prompt, an agent is a model that calls tools (files, browser, APIs, shell, payments) and continues until a stop condition. A chatbot with plugins you click yourself is still chat.
Families you will actually meet
Computer-use / operator-style assistants. Demos where the model clicks a browser or desktop. Spectacular when they work. Fragile on messy UIs, logins, CAPTCHAs, and anything that looks like a bank. Treat as draft-level until you have a trace and a fence.
IDE / repo agents. Coding assistants that edit trees, run tests, open PRs. This is the most mature agent-shaped workflow for many teams because tests are a critic. Still: no unsupervised merge to production.
Support and ticket agents. Fine for classification and drafts. Dangerous if they send without a human on refunds, legal, or medical.
Commerce agents. Shopping and paying on your behalf. The rails got real names in 2025–2026 (assistant checkout experiments, card-network “agent pay” programs, machine-payment protocols). Depth: Pay By Prompt and Buy By Prompt. Do not give spend authority from a vibe.
Ambient / on-device assistants. Summarize, suggest, act on a phone or OS. Often not full agents; they are suggestion engines with a few tools. Ambient vs agent.
We are not listing every startup. Logos rot faster than criteria. Use the scorecard.
How to read a 2026 demo
Ask:
- Sandbox or production?
- Whose credentials?
- Is there a spend cap?
- Can I see the tool trace?
- What happens on a 2FA page?
- Is “memory” files I own, or notes in their cloud?
- What is the retry budget when it gets lost?
If the demo narrator talks only about feelings (“it just handled it”), assume chat with a video editor.
What got easier since 2024
- Tool-calling is normal, not a research paper
- Repo agents + tests are a default for software teams
- People stopped pretending a 2,000-word prompt is a product
- Payment and identity vendors started shipping agent-shaped language (caps, wallets, protocols) instead of only “voice checkout”
What did not magically vanish
- Hallucinated URLs and prices
- UI agents stalling on auth
- Secret leakage into logs
- Unsupervised send/spend incidents
- The need for context packs
Prompting did not die. It got demoted. When you still type.
How we will update this page
When a named capability is widely available or widely retracted, we revise the family descriptions — not a hype ranking. Corrections: Contact. Methodology: methodology.