Products and agents
Agents
Five agents, each with an identity that outlasts any model, a stated list of what it may do, and receipts for what it did.
i
A small team of agents you run, each with its own job, working from one shared memory instead of one bot doing everything.
i
How Guth is built: one identity and one memory for every app, agent and job, shared across apps, devices and models. It lives in AgentNotebook. Private preview.
The five agents
Personal agent: Guth
Your all-purpose personal assistant. Guth carries everyday requests through: calendar, tasks, deadlines, notes, files and research. It handles ordinary work directly, without handing you off to a different persona for each small job.
Coding and security agent: Stan
Stan works on code, reviews changes, runs tests, prepares releases and looks after accounts, permissions and data protection. Stan's app is StanEngineer.
Research and experiment agent: Tilly
Tilly helps turn a question into an experiment, compare models and settings, read the results, and improve how agents behave. Tilly works in Labs.
Business agent: Atlas (GuthBusiness)
Atlas will help with planning, operations and analysis for a business. It is coming soon Planned.
Social agent: SocialCurator
SocialCurator helps you plan, draft, review and schedule posts for Instagram, Facebook, TikTok and X. It is in private preview Private preview; publishing straight to those platforms is not open yet.
Behind these five, smaller helpers do focused background work. They support the main agents and do not need names of their own.
An agent is not a model
i
Who an agent is and how it works: its name, role, personality and instructions. It stays the same when the AI model behind it changes.
i
The AI system doing the thinking for an agent at a given moment. Models can be swapped; the agent keeps its identity, history and knowledge.
Guth keeps these things apart, because mixing them up causes mistakes:
- Identity and instructions: who the agent is and how it should work.
- The model: the AI currently doing the thinking.
- Tools and permissions: what the agent can reach in the place it is running.
- History and knowledge: what it has saved over time.
- Current context: what it actually has in front of it for this task.
A familiar name on its own does not prove continuity. An agent counts as the same agent only when it is connected to its own saved identity, not just because it uses the name.
Permissions: what an agent may do
i
The stated list of what an agent may do, such as read notes or draft a message. Anything outside the list is not allowed.
- You approve consequential actionsbefore they happen.
i
Your yes or no before an agent takes an action with consequences, like sending, publishing, paying or deleting. It shows what will happen first.
- The approval shows what will happen, what will be sent, and to whom.
- Saying yes once does not mean yes forever.
Receipts: what an agent did
i
A record an agent leaves after an action: what it did, when, with what, and what came of it. It lets you check work instead of trusting a summary.
Where to go next
- AgentNotebook: where each agent keeps its identity, notes and history.
- Privacy and security: approvals and data in more detail.
- Release status: what each part can do today.