Agentify · Active R&D

Keeping agent state and control in the application.

Agentify explores how tool-using AI agents can live inside a conventional application where tasks, messages, tools, retrieval, model choice, failures, and human intervention remain explicit parts of the product.

Active R&D

The problem

A useful agent needs more than a model loop.

A useful agent needs more than a prompt and a model response. It needs durable state, authorized tools, retrieval, messages, failure handling, context limits, and clear places for people to intervene.

If those concerns disappear inside a model-centric runtime, the surrounding product becomes harder to understand, operate, and govern. Agentify tests the opposite boundary: the application remains the system of record.

Key decision

Models are replaceable components. The application owns task state and control paths.

Agentify is a Rails application with persisted Task lifecycle state, messages, and tool calls. Model selection is resolved through an account-level gateway rather than embedded in workflow state.

Authorized tools are assembled from account and agent configuration plus system, human, and knowledge capabilities. Context limits, asynchronous processing, errors, and terminal task states remain ordinary application concerns.

Architecture

The system around the decision.

  1. 01Rails task + message state
  2. 02Account-level model gateway
  3. 03Authorized tool definitions
  4. 04Knowledge / retrieval layer
  5. 05Failure + human control paths

What this demonstrates

What this work demonstrates.

Agentify shows how Amidship approaches agentic software when it needs to behave like real software: explicit state, bounded capabilities, replaceable models, observable failures, and human control. The objective is not maximum autonomy; it is an operable product that can use autonomy where it earns its place.

  • Agent architecture inside a conventional product stack
  • Provider and model abstraction without giving the model ownership of workflow state
  • Persisted task, message, and tool-call behavior
  • Retrieval, integrations, and human-control concepts as application capabilities
  • Operational handling of context limits, asynchronous work, errors, and recovery

Inspectable evidence

Inspect it directly.

Open Agentify (opens in a new tab)

Evidence boundary

Agentify is active R&D, not a client deployment. Current domain, infrastructure, and front-desk product work remain under development and are not represented as shipped production capability.