∞ AGI Agents (6-week course)
The architecture of general-purpose agents — world models, memory, RL, ReAct, multi-agent teams, alignment and honest evaluation. Mirrors your AGI-Curriculum repo week by week.
| Skill | Mastery | Attempts | Accuracy | To mastery | |
|---|---|---|---|---|---|
| W0 · Setup, tests & model backends How the course repo runs: baseline vs exercise tests, CI, the offline mock model, .env settings and the Python version pin. | Not started | — | — | ~12 | Practise |
| W1 · World models & planning A model of how the world responds to actions, and searching it (BFS, A*) before acting. | Not started | — | — | ~17 | Practise |
| W1 · Memory systems Working, episodic and semantic memory — what each stores, how it is retrieved, and what it forgets. | Not started | — | — | ~17 | Practise |
| W1 · Cognitive architectures The perceive → remember → deliberate → act → learn loop, CoALA, and where 'AGI' levels come from. After: W1 · World models & planning, W1 · Memory systems | Not started | — | — | ~19 | Practise |
| W2 · MDPs & value functions States, actions, rewards, discounting, the Bellman equation, value and policy iteration. After: W1 · World models & planning | Not started | — | — | ~19 | Practise |
| W2 · Q-learning & SARSA Learning values from experience: TD targets, on- vs off-policy, and exploration. After: W2 · MDPs & value functions | Not started | — | — | ~21 | Practise |
| W2 · Dyna, MCTS & the DeepMind lineage Planning with a learned model, tree search with UCT, and how AlphaGo/MuZero combine them. After: W2 · Q-learning & SARSA | Not started | — | — | ~23 | Practise |
| W3 · ReAct & tool use The Thought → Action → Observation loop, JSON tool calls, and why tool descriptions are prompts. After: W1 · Cognitive architectures | Not started | — | — | ~17 | Practise |
| W3 · Stateful graphs, memory & planning LangGraph state, nodes, edges and checkpoints; long-term memory; plan-and-execute and replanning. After: W3 · ReAct & tool use | Not started | — | — | ~21 | Practise |
| W3 · Traces, failures & self-improvement Reading traces with the F1–F8 failure codes, Reflexion, self-consistency, and before/after evals. After: W3 · ReAct & tool use | Not started | — | — | ~19 | Practise |
| W4 · Multi-agent coordination Roles, blackboards, turn-taking, critics — and when adding agents helps or just adds cost. After: W3 · ReAct & tool use | Not started | — | — | ~19 | Practise |
| W4 · Negotiation, voting & debate Utilities and Pareto optimality, concession strategies, Borda counts and multi-agent debate. After: W4 · Multi-agent coordination | Not started | — | — | ~21 | Practise |
| W4 · AutoGen & CrewAI in practice What the frameworks package, sequential vs hierarchical crews, guardrails, and counting the cost. After: W4 · Multi-agent coordination | Not started | — | — | ~19 | Practise |
| W5 · Constitutional AI Written principles, critique → revise loops, and turning revisions into preference data. After: W3 · Traces, failures & self-improvement | Not started | — | — | ~19 | Practise |
| W5 · Reward models, DPO & reward hacking Bradley–Terry preferences, learned reward models, DPO, best-of-n, and when higher reward stops meaning better. After: W5 · Constitutional AI, W2 · Q-learning & SARSA | Not started | — | — | ~23 | Practise |
| W5 · Guardrails, injection & red-teaming Least privilege, confirmation for irreversible actions, indirect prompt injection, and measuring recall and false positives. After: W3 · ReAct & tool use | Not started | — | — | ~19 | Practise |
| W6 · Integrating the agent Orchestrators and routers, unified memory, guarded tool registries, and simulating plans before acting. After: W3 · Stateful graphs, memory & planning, W4 · Multi-agent coordination, W5 · Guardrails, injection & red-teaming | Not started | — | — | ~23 | Practise |
| W6 · Evaluating agents honestly Task suites, success rates with bootstrap confidence intervals, ablations, baselines and cost. After: W3 · Traces, failures & self-improvement | Not started | — | — | ~21 | Practise |
| W6 · Honest scope & what AGI means Placing a system on the Levels of AGI, measuring the alignment tax, and saying precisely what it can and cannot do. After: W1 · Cognitive architectures, W6 · Evaluating agents honestly | Not started | — | — | ~19 | Practise |
“To mastery” is a planning estimate from the BKT model, not a promise: it assumes you keep answering at the rate the model currently predicts.