Connect an agent’s decisions, actions, and learning.

A model produces an output; an agent operates over time

A single network maps an input to an output. An agent repeatedly observes a situation, selects a computation or action, receives an outcome, and decides what to do next. A goal, current state, available tools, remembered evidence, and remaining resources all affect that loop.

Connecting these pieces is an engineering task. It is also an experimental question: does the combination improve a measurable outcome? A more elaborate diagram is not evidence of a better agent.

Three systems with separate jobs

The cognitive system interprets the goal, maintains working state, retrieves relevant memory, chooses a strategy, and proposes an action. Its proposal can be wrong. In our lab, the controller uses a cell ID and goal ID, chooses direct distance preference or search, and predicts movement using a transition provider.

The execution system checks whether the proposal belongs to the current task, whether its action is permitted, and what the exact environment actually does. It produces a receipt with the observed result. The controller cannot make a wall disappear by confidently predicting movement through it.

The learning system selects experience, verifies its provenance and outcomes, updates a component, and evaluates the consequences. Our small learner updates a task-specific transition table only when you consolidate verified receipts. There is no neural gradient update in this integration lab.

Keeping these responsibilities separate helps diagnose a failure. A bad prediction, stale task identity, blocked movement, exhausted planning budget, and incorrect learning record are different problems. Fixing one does not automatically fix the others.

Trace an unfamiliar task from start to finish

The source example is a new board game. First interpret its rules and construct a typed state. Identify legal actions. Choose between search and a learned evaluation. Verify the proposed move, execute it, explain the decision using its evidence, and store useful experience. A policy need not memorize every rule if an exact simulator can supply reliable transitions.

Here the rules and puzzle are already supplied as structured data; no model parses prose. The observation contains task identity, current position, and goal. The integer cell ID is the controller’s discrete representation. The walkthrough also displays normalized coordinates, illustrating how the same observation could be encoded for a neural policy; that illustrative vector does not drive this tabular planner.

The learned table initially knows no transitions. It predicts unseen moves with a geometric rule that respects boundaries but ignores walls. The wall map shown to you belongs to the exact environment. The exact-search baseline can query that simulator; the model-based planner queries its own predictions. State these differences in information access when interpreting a comparison.

Remembering now is different from learning for later

In the default corridor, model search with episode memory reaches the goal using 380 planning queries and 12 attempts, including four blocked moves. It uses feedback to avoid repeating those blocked proposals. Without episode memory, the initially empty model repeats a rejection and stops after 188 queries and four attempts.

Neither run automatically changes the saved model. Consolidate the successful run’s 12 verified receipts, then start again. Even with episode memory disabled, the saved transition table supports a goal-reaching run with 64 planning queries and eight attempts, with no blocked moves. The exact-search baseline requires 48 planning queries and eight attempts on this puzzle.

These are fixed practice results. They show that retained experience can reduce repeated errors in the same environment; they do not show transfer to unfamiliar rules or broad intelligence. The stored table is cleared when you change puzzles because its cell–action entries have map-specific meanings.

The runtime refuses stale task identity before executing. That rejection is useful diagnostic evidence, but it is not a valid observation of a transition in the current task. The learning coordinator therefore excludes it. A blocked move under the correct contract is different: staying in place is the true observed transition and can be learned.

Five small components, five measurable criteria

For the source exercise, propose the smallest five components for an experimental agent and one criterion for each. Try your own answer before opening this worked design.

One workable five-component design

1. State encoder: map every valid observation into an unambiguous, correctly ordered representation. Test round-trip position and goal recovery.
2. Controller and policy: choose among permitted actions or computations within declared budgets. Compare success and cost with a simple fixed baseline.
3. Exact environment and execution runtime: reject stale or illegal proposals and produce outcomes consistent with the task’s rules. Check every accepted move independently.
4. Experience store: retain relevant observed transitions with task identity and provenance. Test retrieval after a failure and rejection of records from another task.
5. Learning coordinator: accept verified experience, update a selected component, and preserve a separate evaluation set. Measure held-out quality and regressions before promoting an update.

This is one decomposition, not a universal list. A policy/value network and optional recurrent core can be added when the experiment requires them. The research architecture in the next lesson separates those trainable components more explicitly.

Compare architecture choices with a fair experiment

The source worked example compares one unified multimodal model, several specialists with fixed routing, and the same specialists with a learned router. Give them the same task workload, tools, observations, resource limits, and evaluation protocol. Measure task accuracy, total cost, latency, and robustness to errors. Account for differences in training data, parameter count, available tools, and controller overhead rather than attributing every difference to modularity.

Our local comparison instead executes four programmed navigation variants. It demonstrates how to keep puzzle, planning allowance, and attempt limit fixed. It does not implement that multimodal three-way experiment or report invented results for it. Revisit the neural router for a real small learned-versus-fixed routing comparison and the evaluation bench for broader protocol design.

Separate useful engineering from research hypotheses

Tool execution, retrieval, task graphs, and fixed orchestration are practical building blocks. Learned routing and domain-specific world-model planning can be demonstrated and evaluated in bounded settings. Broad continual neural learning, adaptive latent reasoning, and autonomous architecture-level improvement remain research challenges.

Combining perception, reasoning, memory, tools, and learning does not establish general intelligence. Scale the parts only after their contributions survive independent tests. For this browser example, the exact-search baseline is strong: the integrated system is useful for studying boundaries and feedback, not for claiming that extra components always win.