Skip to content
dont dumb
DOCS / ARCHITECTURE

One host. Two responsibilities.

Think of DontDumb as a learning subagent with its own harness, operating alongside the main agent through the same host connection.

YOUR EXISTING AI CONNECTIONSeparate learning work · additional tokens
01 / DELIVERY

Coding agent

Owns the task.

  1. Your task
  2. Plan → implement → test
  3. Working software

Keeps working independently.

Selected
work events
02 / UNDERSTANDING

Learning harness

Owns the learning loop.

  1. Notice a relevant idea
  2. Invite → explain → practice
  3. Assess → adapt ↺

Local learner context
Goals · knowledge · practice evidence

Context guides the next opportunity.

Selected context crosses the boundary. Teaching responsibility stays with DontDumb.

What belongs to each loop

The coding agent plans, implements and tests. The learning harness observes permitted events, selects meaningful context, considers learner goals, offers practice, assesses evidence and updates memory. Coding does not wait for a lesson or an answer.

What “within your host” means

The existing host supplies model execution and account access. DontDumb supplies a separate local learning runtime. Codex uses separate model work through its app-server; Claude’s local module bridges the selected Code session; OpenCode provides a separate learning session. The role is a subagent, but the exact mechanism is not identical across hosts.

The local companion server is an on-device process serving your panel and coordinating local state. It is not a remote DontDumb backend. We do not run a DontDumb model service.

Three core inputs

  • Worker context: bounded selected task events, captured changes and tool outcomes.
  • Learner context: explicit goals/preferences and evidence of practice.
  • Learning tools: explanations, questions, maps, interactives and assessments.

Execution boundary

The model proposes tool actions. The runtime validates inputs/results, applies bounds and cancellation, and owns state commits. Work selection avoids forwarding every output token. The adapter must not assume full repository access or hidden reasoning.

Learning calls use additional resources and can share model rate limits with coding. Independent execution means no learning gate in the coding task; it does not promise zero throughput impact.

Implementation map

packages/runtime/src/learning-engine.js  orchestration and state
packages/runtime/src/agent-cycle.js      bounded agent cycle
packages/runtime/src/tool-registry.js    registered tool contracts
packages/learning/src/work-selection.js  work sampling
packages/learning/src/learner-store.js   local shared learner state
adapters/codex · claude · opencode       host integration