CORE is a cognitive harness paradigm for coding agents, developed as part of the broader open source research at Ori Mnemos on modeling human cognition in language model systems in a manner similar to how a plane models a bird.

The Core Thesis

Agents should not reason through bash.

Bash is an unstructured string interface with no object models, no composable namespace, and no capacity for recursive dispatch.

Language models reason natively through structured token generation — composing expressions, operating on typed objects, building outputs from intermediate results. Bash breaks this at every serialization boundary; a REPL environment preserves it richly.

CORE transforms the codebase into a knowledge graph and exposes that graph as a Python REPL environment the agent can natively traverse.

A New Execution Environment

Inside this environment, the agent writes Python that composes operations in a single turn:

  • search the graph
  • cluster results by file
  • fan out to fresh LLM sub-reasoners per cluster via rlm_call and rlm_batch
  • synthesize the outputs

One expression replaces what tool-calling architectures require ten or more sequential round-trips to accomplish.

Recursion as the Primitive

These recursive language model calls are the core primitive.

A language model, mid-reasoning, spawning focused instances of itself on decomposed sub-problems and composing the results back into a unified output.

Recursion applied to inference.

Why This Outperforms Bash

Bash requires the agent to:

  • serialize every operation into a string command
  • wait for unstructured output
  • parse that output back into something meaningful
  • decide the next command

Each step is a lossy boundary.

Structure is destroyed on the way out and reconstructed — if at all — on the way back in.

At scale:

  • ten sequential tool calls
  • ten serialization boundaries
  • ten parse attempts
  • ten opportunities to misinterpret output or lose context

Eliminating the Boundary

CORE eliminates these boundaries entirely.

The agent operates on structured objects:

  • graph nodes with PageRank scores
  • clustered search results
  • typed function returns