Bot landscape | hotbotspot.com

What we mean by a bot

Everyone calls everything a bot now: chat windows, coding agents, phone assistants. Here's the definition we use. A bot is a persistent agentic model that is durable and teachable. It keeps its identity and purpose while maintaining a set of memories, skills and playbooks. Tools help it get work done, but they're not part of the definition.

A one-pass agentic system finishes a task and stops. A long session compacts as it goes, and a rule given at the start can be filtered out. A bot keeps its identity, standing rules and playbooks outside the chat and reloads them every time, so its core survives compaction.

The test is continuity. Close the chat and come back tomorrow. Does it still know who it is, what it's for, and what you taught it?

One correction up front. Bots don't do away with context limits, and they don't necessarily remember everything. What they add are ways to keep and recover information outside the current conversation. Each product does this differently. docs.x

	Bot	One-pass agent
Job	Ongoing	Single task
Memory	Kept outside the chat	Ends with the task
Workspace	Survives sessions	Temporary
Next assignment	Uses what it learned	Starts fresh

Agent vs. bot

An agent takes a goal and acts on it. A bot is one kind of agent: one that lasts.

The difference shows up over time. A long agent session doesn't run forever on raw history. It compacts. It summarizes old turns and throws away detail so it can keep going.

Each compaction is a filter. Most of what it drops is noise. But nothing stops it from dropping something important too. Say you told it at the start, "don't set the building on fire." After enough rounds, that can get filtered out like any other detail. Now you're working with a different agent than the one you started with, and it doesn't know the one rule that mattered.

A bot keeps its identity, purpose, standing rules and playbooks somewhere separate from the chat, and reloads them every time. The conversation can still compact. The core of the bot doesn't.

A bot notices what to keep

In a normal agent session, you have to spot the moment something matters and say "write this down."

A bot is supposed to watch for that itself. Correct one, and a good bot asks some version of: "Should this be the new default?" That question is the bot deciding whether your correction belongs in its standing memory or just in today's chat.

Three groups

Group	Products	What you're really choosing
Continuing work assistants	Grok Bot, OpenClaw, OpenAI dots, Meta Muse, Claude Duet, Perplexity Computer	An assistant's memory controls, integrations, permissions and background behavior. OpenClaw is self-operated: you run your own workspace, memory files, channels and automations.
Persistent engineering and coding workers	Cursor Automations and Cloud Agents, Claude Code, Grok Build, OpenCode	How an agent does the work, keeps project instructions and carries learning across sessions.
Foundations to build your own	Claude Agent SDK, Claude Managed Agents	The building blocks, for when custom development makes sense.

Sources: docs.x, help.openai, meta, claude, docs.openclaw, openclaw, code.claude, cursor, x, code.claude SDK, platform.claude

Which to choose

Assessment

These are architectural fits based on the documentation. They are our assessment, not a claim that any product is better across the board.

If you want…	Start by looking at	The caveat to check
A team of named specialists that hand off work and keep their own role context	Grok Bot	All bots share one computer. Separate names are not separate security compartments.
To study memory you can see and edit as files	Meta Muse	Deleting something doesn't necessarily remove everything learned from it.
A proactive professional assistant built on Claude skills you already have	Claude Duet	Preview access and shared usage limits may make testing harder.
Continuity with ChatGPT and ongoing work across apps	OpenAI dots	You can't view or edit individual dot memories.
A continuing coordinator that you run yourself	OpenClaw	You operate and configure it.
Recurring work that checks what changed since the last run	Perplexity Computer with Projects, Brain and Automations	Learned context isn't an approved decision log.
Recurring engineering or software jobs with memory you can edit	Cursor Automations	Memory belongs to one automation, and untrusted input can plant bad memories.
A project worker that keeps instructions and learns from corrections	Claude Code, Grok Build	Having memory doesn't make it an always-running coordinator.
A project worker you can customize and extend	OpenCode	Compare a specific setup, including its plugins.
To build your own product	Claude Agent SDK, Claude Managed Agents	These are building blocks, not a ready-made assistant.

Sources: docs.x, meta, claude, openai, openclaw, perplexity, cursor, code.claude, x, github, code.claude SDK

One thing holds for every row: no bot's private learned memory should be the only record of a project. See Memory and testing.

A caveat about Grok Bot

All the bots on one Grok account share one cloud computer, including its files, browser sessions and app logins. Each bot gets its own screen for parallel work, but those screens are not separate accounts or machines. That makes it easy for bots to cooperate on one business's projects. On its own, it isn't enough to keep confidential clients apart. docs.x

Bot technologies: the landscape