Intelligence in Depth layers

Intelligence in Depth (IID) organizes model-assisted systems into six layers. The layers are composable: use the shallowest sequence that reliably satisfies the task. See …But Is It Agentic? for the architecture’s principles.

1. One-time prompting

An application sends a direct instruction to a model through a library or API. This layer suits a single transformation or answer whose inputs and outputs are already defined.

For example, an application can send one research narrative to acorn-lib’s OpenAI-compatible chat-completion API with the instruction “Summarize this activity,” then return the response.

“OpenAI-compatible” describes the connection format, not the model provider. A compatible server accepts supported OpenAI-style HTTP endpoints and JSON requests and responses. The server may be OpenAI’s service or another implementation, including llama-swap, vLLM, or LiteLLM. ACORN uses this format because local and hosted servers widely support it, which reduces custom client code.

2. Prompt templates

Prompt templates store reusable instructions in version-controlled .prompt files with defined inputs, configuration, and expected outputs. Google’s Dotprompt specification documents this executable file pattern.¹

For example, ACORN renders its embedded summarize.prompt template with a source narrative, maximum word count, and output structure so callers can reuse the same instructions.

3. Workflow composition and iteration

A workflow coordinates model calls, deterministic operations, and tools through explicit control flow. It may sequence, select, repeat, delegate, or combine bounded tasks, but the application defines its path and safeguards. Anthropic distinguishes these predefined workflows from agents that direct their own process and tool use.²

Repeated workflows need clear exit conditions, measurable progress, state passed between calls when required, and limits on iterations, time, tokens, and cost.

A planned ACORN workflow could ask a model to address gaps in a research narrative, run deterministic validation, and pass remaining findings to another bounded attempt. It would stop when no high-severity issue remains or when it reaches an iteration, token, or time limit.

4. Skills, context, and capabilities

This layer assembles the instructions, research data, evidence, examples, tools, constraints, and current state needed for an invocation. Agent Skills package specialized instructions and optional resources into portable, version-controlled directories.³ MCP can connect an agent to live prompts, resources, and tools.⁴

This is the layer where an application makes a model context-aware for a specific task. The model can use supplied relationships and provenance to explain connections or identify gaps, but it cannot treat an unsupported inference as source data.

For example, an agent can load the ACORN skill with a ResearchActivity schema, source record, validation results, and allowed tools before proposing a correction. The local stdio MCP server exposes the same canonical tool registry and deterministic checks through a standard interface.

5. Connected memory

Connected memory preserves selected state across iterations, sessions, or tasks. Useful memory includes attributable facts, decisions, provenance, and prior results. Its origin, availability, modality, and quality are part of its data profile.

A planned workflow could retrieve a previously reviewed project identity, identifier mapping, and provenance record instead of reconstructing them from conversation history.

6. Agent harness and interoperability

An agent harness orchestrates models, skills, tools, workflows, context, memory, permissions, validation, recovery, observability, and human approval. It owns resource budgets and stopping conditions. Permissions and approvals shape autonomy, while perception, state, actions, and feedback shape motivity. ACP can standardize sessions, streamed updates, and permission requests between a harness and external agents.⁵

ACORN launches OpenCode through ACP for bounded fix suggestions, sends the normalized check results as context, streams progress, and applies a fixed permission policy. The returned suggestions remain reviewable data; acorn check does not apply them.

ACORN’s Protobuf/gRPC transport provides part of the foundation for that work. Its shared operation registry, policy context, authentication, deadlines, and cancellation behavior can support more protocol adapters. ACP will still use ACP sessions, streaming, and permission messages, while MCP keeps its own tool discovery and invocation semantics.

Applying the principles

Current ACORN support

ACORN can develop each layer independently. Prompt evaluation can precede workflow execution, deterministic operations can become tools before broader agency is granted, and memory provenance can be defined before agent state is retained.

1. Google, Dotprompt specification and Dotprompt reference documentation. ↩

2. Anthropic, “Building effective agents”, 2024. ↩

3. Agent Skills, Specification. ↩

4. Model Context Protocol, “Architecture” and “Server features”. ↩

5. Agent Client Protocol, “Architecture”. ↩