Who it is
Purpose, reporting line, values, tone, and non-negotiable boundaries.
A Guth Agent has a readable charter, bounded jobs, explicit authority, versioned instructions, source-linked memory routes, field guides, training pairs, and a receipt history. The identity can move across applications without pretending that every new model session remembers who it is.
Public profiles disclose roles and boundaries, not secrets, private memory, credentials, or raw conversation history.
The profile tells a model how to become the same Agent again under controlled conditions. It points to current evidence rather than stuffing an entire history into every prompt. Each component has its own owner and revision state, allowing a charter to remain stable while field guides, projects, or active work change.
Purpose, reporting line, values, tone, and non-negotiable boundaries.
Named responsibilities, expected outputs, and the work it should actively maintain.
Allowed local actions, effects requiring a principal, and actions it will never take.
Current rows, packets, corpora, projects, receipts, and approved source partitions.
APIs, MCP tools, skills, instruments, documentation indexes, and working patterns.
Curated input-output examples plus rejection and repair evidence for difficult cases.
Profile versions, approvals, predecessors, successors, and the reason for each revision.
Source-bound records of work, holds, reviews, handoffs, effects, and recoveries.
A public profile should make an Agent understandable without exposing private state. Select a profile to inspect its jobs, principal-required actions, and permanent limits. These cards describe role packages; an active session must still load and verify the current profile version.
In the documented model, the interface changes with the job: Notebook is natural for notes and tasks, Files is natural for browsing and recovery, Calendar is natural for time, and Labs is natural for experiments. The model keeps one named Agent identity across surfaces while each verified application receives only the context and tools needed there. This page does not demonstrate cross-surface enforcement.
The intended Pointguard would verify the Agent profile, bind current scope, retrieve approved context, preserve receipts, and prevent one surface from silently manufacturing authority for another. Those controls require separate implementation evidence.
Choose an Agent, surface, and task scope. The activation preview shows the package loaded for that session. A real activation would resolve signed profile versions and current database records; this public demonstration is local and uses no private corpus.
Local illustration only. No profile registry, signature, database, private corpus, or receipt service is queried by this page.
An Agent can learn a new platform, receive a new job, lose a permission, or be succeeded by another Agent without pretending the change never happened. Guth versions the profile and records who approved the transition. Prior versions remain inspectable evidence rather than hidden prompt history.
Naming and adoption require explicit human review. Fine-tuning, repeated use, or model confidence does not independently create an Agent identity.
Persistent identity depends on reliable retrieval, not an unlimited transcript. The profile names the authorized memory partitions and retrieval rules. Current work is loaded first; older evidence is retrieved only for the concrete question. Training pairs and rejection cases help reproduce quality without granting remembered text new authority.
An Agent coin is a human-readable issuance object, not a magical credential. Its struck appearance can be unique and imperfect, while the underlying Agent Stamp binds the visible object to an issuance receipt, profile digest, organization, and version. Verification depends on cryptographic records, not on visually recognizing the coin.
The strike can vary in centering, pressure, wear, and surface detail. The model would bind an issued object to a receipt and digest that can be checked without trusting an image. No verification record is queried or demonstrated on this page.
Illustrative flow: profile version → issuance receipt → SHA-256 digest → verification recordThese terms separate identity, memory, authority, and model execution so one cannot quietly substitute for another.
Loading a verified Agent profile, current scope, and approved context into a bounded working session.
A named, versioned role package with jobs, boundaries, routes, and accountable continuity.
The durable statement of an Agent’s purpose, reporting line, values, and non-negotiable limits.
A principal granting limited authority for a specific purpose, scope, and duration.
A source-backed package of tools, documentation, patterns, and instruments for a specialty.
A reviewed example connecting an input with an accepted output and the reasoning or evidence behind it.
The inspectable chain linking an Agent’s origin, profile versions, approvals, and successors.
The continuity layer that verifies identity, binds scope, retrieves context, and preserves receipts across surfaces.
A cryptographic hash used to identify an exact profile version without exposing its private contents.