Skip to main content
Version: Next

Three-layer Architecture

1. Why Prodia Agent is better understood through three layers

Many teams initially picture an Agent as “an LLM, a prompt, and some tools”. That model is fine for lightweight demos, but it is not enough for long-running industrial use.

Manufacturing analysis is different because:

  • problems are multi-turn rather than one-shot
  • results must be grounded in real industrial capability and real data
  • the system has to decide not only what to say, but also what to see and what it is allowed to do

That is why the current Prodia Agent is better understood through a three-layer architecture.

2. What the three layers are

LayerCore questionRole in Prodia
Harness EngineeringWhat execution world does the Agent run inside?provides sessions, tools, state, events, permissions, and runtime governance
Context EngineeringWhat is this turn actually allowed to see?decides history, memory, knowledge, time, tool boundary, and token budget
Prompt EngineeringHow should this turn behave and respond?defines role, rules, mode switch, output contract, and subtask boundary

In simple terms:

  • Harness builds the world
  • Context selects what the turn sees
  • Prompt shapes how the turn behaves

3. Why prompt alone is not enough

If prompt is expected to carry all the work by itself, several problems appear:

  • history grows, but signal quality drops
  • tool boundary and execution order become harder to stabilize
  • one large prompt becomes responsible for role, process, context, safety, and governance all at once
  • the system starts depending too much on ad hoc model behavior

That is why Prodia is no longer best framed as “prompt-first”. It is better framed as runtime plus context plus prompt working together.

4. How the three layers collaborate

More concretely:

  1. Harness prepares the real execution environment: session, tool access, state, and event path.
  2. Context decides which history, memory, glossary, knowledge, and time scope should enter the turn.
  3. Prompt turns those inputs into stable role behavior, output discipline, and response boundaries.

These layers do not replace one another. They complete one another.

5. What product value the three layers create

LayerDirect product value
Harnessmakes Agent executable rather than merely conversational
Contextmakes Agent selective and relevant rather than overloaded or forgetful
Promptmakes Agent behavior stable rather than loosely improvised

6. How the three layers relate to other Prodia mechanisms

The three-layer architecture does not replace UNS, MCP, Skills, memory, or the knowledge loop. It gives them a clearer technical place:

  • UNS, MCP, and tools sit closer to the Harness side
  • routing, context engineering, and memory sit closer to the Context side
  • prompt assets, output contracts, and subtask constraints sit closer to the Prompt side

This makes the platform easier to extend and easier to govern as scenarios grow.

If you want to go deeper into what the Agent sees and how it behaves, continue with Context Engineering and Prompt Engineering.