Bounded agents, by declaration and composition

TAP is the Thinkun Agent Platform: a platform for defining agents as reusable AI Agent Products, composing them from reviewed Facilities, and deploying them into an organisation with explicit authority, chosen model inference providers and a durable record of work.

A specialist focused on a task, an agent per staff member, or a whole team of digital co-workers.

Discuss TAPHow an agent gets made

The bounded AI agent is the product. The boundary is the asset.

A different starting point

Most agent platforms optimise for breadth: more tools, integrations, a marketplace. TAP, built by Thinkun a quarter century into software and operations for business, starts from the job and its boundary.

A general agent

  • Every tool it can reach
  • Permissions negotiated in the prompt
  • Restricted to one AI lab's models
  • A chat history
  • Skills pulled from a marketplace

A TAP agent

  • A defined job and the Facilities that job needs
  • Authority granted by people and enforced at execution
  • Hundreds of hot-swap models for AI inference
  • Durable sessions, records and provenance
  • Compositions reviewed before they can be installed

Agent Products

TAP defines Agent Products as reusable, composable declarations of a kind of agent: its identity, operating instructions, specialist skills and the Capabilities, the work, it offers. TWA is an Agent Product; your webmaster is an agent deployed from it.

One declaration, many agents

The product carries the meaning. A deployment supplies what differs: people, systems, providers, limits and separately granted authority.

Composed, not coded

A product composes the Facilities its Capabilities need. New code is written only for a missing building block, and it extends the platform, not one agent.

From declaration to running agent

Resolution is validation, not convention: an unknown component, missing provider feature or invalid binding refuses the composition before the agent exists. What passes is frozen, with a stable identity.

  1. Agent Product

    identity, persona, operating instructions, product skills, base Facilities and offered Capabilities.

  2. Capability

    a complete outcome a customer can recognise and choose. It composes Facilities, guidance and skills; it never confers permission.

  3. Facility

    a reviewed technical composition: tools, skills, integrations, required provider features, eligible kinds of authority.

  4. Providers

    one implementation behind each generic contract: inference, execution, channels, source, delivery, state, web access.

  5. Deployment

    one installation: product, people, provider and resource bindings, constraints, credential references, selected Capabilities.

  6. Resolved and run

    TAP joins the declarations against the installed platform, refuses anything incoherent, freezes and mounts the result.

Ability is not authority

Composition, visibility, entitlement and role never become permission. Only a grant says this principal may take this action on this resource, now.

Explicit grants

A grant names the principal, the enabling Facility, the resource, the action and its constraints, and is checked at execution, outside the model's conversation.

Monotonic denial

Policies can narrow or ask; nothing can widen. Missing, expired, malformed or replayed authority denies. Misconfiguration never opens a boundary.

Evidence before irreversible actions

TWA prepares a change in isolated execution, stages it with server-side evidence, and can promote only that exact reviewed release under a separate human grant.

Records and recovery

Durable records say what happened. An uncertain external effect is reconciled against authoritative state, never blindly retried. Upgrades are pinned and reversible.

Providers are bindings, not the product

Each deployment binds the providers it needs, so the agent, its rules and its authority belong to no lab, model or account.

Cost is a binding too

Each deployment carries its own inference bindings and spend limits. A provider cap becomes a plain message, not a fallback to another account.

TWA: Agent Product spotlight

Thinkun Webmaster Agent declares one base Capability, web development, composed from four Facilities: web authoring, source control, web delivery and release promotion. A deployment binds it to a site, people and providers. TWA stages a change; only a person can promote the exact release they reviewed.

TWA for website owners, at Thinkun

Built on TAP

TWA

Website stewardship and small web apps, by conversation.

Meet TWA

TOM

Operations co-workers defined for one business.

Meet TOM

Mullum

Life and Work Agents, each a separate deployment with its own identity and boundaries.

Visit mullum.ai

Is it no-code?

No. Where TAP has the building blocks, a new agent is mostly declaration, composition and review. A missing tool, integration or adapter needs engineering, then becomes reusable.

Can I put my own agents on it?

TAP runs Thinkun's Agent Products today. For an agent per staff member, a team of digital co-workers or an Agent Product of your own, tell us.

Discuss TAP

Tell us what you have in mind: an Agent Product, digital co-workers, or AI software development services.

What would you like to explore?