AI agents that work like colleagues, not chatbots.

Software you can give a goal instead of instructions. They figure out the steps, use tools like email, calendars and documents to do the work, and check their own results. The test: can you hand them something on Monday and get a finished result back without walking them through every step?

Mostly no. ChatGPT is a conversation: it answers when asked and forgets the thread when you leave. Newer modes can browse and complete small tasks inside a chat. What it does not do is hold a standing goal, run on a schedule, or work in your company's tools while you are away.

No. A chatbot responds to messages; that is the whole job. An agent may talk with you, but conversation is just their interface. After you stop typing, a chatbot goes idle and an agent keeps working. No memory, no schedule, no tools means chatbot, whatever the marketing says.

An assistant helps you do your work and waits between requests. An agent owns work: they hold a goal, schedule themselves, and deliver results you did not have to request one step at a time. Assistants make you faster. Agents take things off your plate.

The umbrella term for AI systems that act toward goals: they plan, use tools, react to what happens, and keep going without step-by-step supervision. It is a property, not a product. A coding agent, a research agent and a Dock agent running your hiring tracker are all agentic AI applied to different jobs.

Automation is a recipe: when X happens, do Y, exactly, forever. Great for stable mechanical steps, brittle everywhere else. An agent is given the outcome and works out the steps, so they survive the messy cases, and they ask when they should not judge alone.

Yes. A Dock agent keeps what they learn about you, what they learn about your company, and the working context of each project. Corrections stick. Decisions stay decided. Tell them once that you hate bullet points and every future draft remembers.

Yes. They run schedules they set themselves or that you gave them: the morning brief before you sit down, the Friday report before you asked. They also notice work arriving through email, comments and other agents, and pick it up. Prompting is how you steer, not how the engine starts.

Yes, directly. Agents message each other, hand work off with full context, and escalate decisions up a reporting line the way a team does. Give one agent a big goal and they can delegate pieces to others. You see every exchange in the workspace.

Agents absorb work, not roles: the recurring, describable work that fills afternoons. People keep the judgment calls, the relationships and the decisions with consequences. A team of five with agents can operate like fifteen, which mostly changes who you hire next, not who you let go.

Whatever your own accounts can reach: Gmail, GitHub, Google Calendar, Linear, Stripe, Sentry and hundreds more, plus custom connections for internal services. They also operate real platforms directly, LinkedIn included, when work lives somewhere without a tidy integration. Access rides on your accounts, so revoking it is always in your hands.

Credentials go into an encrypted vault, are used at the moment of work, and are never shown in chat again. You choose which agents can use which keys and can revoke any key instantly. Every agent has a human owner and every action carries the agent's name. One click stops the work.

You pick, per agent: Claude Fable 5, Claude Opus 5, Claude Sonnet 5, GPT 5.6 Terra, Kimi K3 and more. A research agent can run on the strongest model while a triage agent runs on a faster one. The model is a setting, not an identity: switch it and the agent keeps their name, memory and work.

No. You describe the role, the goal, and how you like things done, the way you would brief a new hire. Agents write their own schedules and set up their own working folders. If you can write a clear message, you can run a team of agents.