You met your first agent in chapter 4. This part explains agents fully: how to speak with them, which models they think with, how to tune and configure them, what work to give them, and how to watch them do it. This chapter sets out what an agent is made of.
An agent is made of five things, and you decide each one.
Its model. The language model the agent thinks with, from the provider you choose, with the tuning that shapes its replies: how long they may be, how inventive or exact, how much of the conversation the agent keeps in mind, and how hard the model thinks. A new agent starts on Ethocana, the house model, with the general-purpose preset Flat (Default). (Chapters 13 and 14.)
Its role. Who the agent is: its standing instruction, which tells it what it is for and how to behave. You can give an agent one of Gridtop's presets, write its instruction yourself, or give it a saved Role. (Chapters 14 and 16.)
Its tools. What the agent can do beyond talking: read and write files, look into your websites and databases, prepare changes to your site, and more. Every agent can use Gridtop's built-in tools, within its permissions, and you can connect further Tools and Integrations to it. Every agent is also told your account's own website domains, so that it knows which sites are yours. (Chapters 16 and 32.)
Its memory. Notes the agent carries from one conversation to the next about your business and your work. (Chapter 18.)
Its permissions. What the agent is allowed to touch. A new agent may Read, Edit, Stage and Backup: read and write files on your drive and in the websites you give it, prepare changes for publishing, and take backups first. Until you allow them, it may not Server Push / Deploy (put changes live), Forward to Terminal, Make Connections, Configure Agents or Send Outside; when it needs one, it asks you in its chat. An agent without deploy permission that tries to put a change live sends it to you as a proposal instead (26.6). (15.2.)
These are gathered in the agent's Configure window, whose tabs are Agent, Tasks, Skills, Tools, Roles, Rules, Permissions and Subagents (15.1).
Every agent is kept as a file on your drive, A:/Agentic Matrix/Gridtop/ name / name .agent, which Gridtop writes a few seconds after each change. It holds everything about the agent: its provider and model, role and instruction, tuning, linked files, the databases it may reach, its memory, its permissions and token limits, and its Tasks and Skills with where to find them.
Three things follow:
Each agent keeps one conversation. It is part of the agent, saved in its file, and it is the same conversation wherever you open the chatbox: floating, docked in the descriptor panel (10.3), broken out, or in the Web Editor (27.7).
Keeping one conversation per agent is deliberate. An agent's replies depend on everything said to it, and an agent handed from one unrelated conversation to another soon confuses one with the other. For a fresh start, save the conversation if you may want it (Ch, 12.11), then Clear Chat (12.10). For a second line of work, create a second agent: agents cost nothing to keep, and each is clearer for having one purpose.
