Skip to content

AI, Autonomous Science & Agent Systems

  • Exit capability

    Build/evaluate agentic systems that can reason, use tools, coordinate, run experiments, and improve workflows without losing independent verification.

  • Frontier targets

    AI scientist; coding agents; research swarms; theorem/reasoning agents; self-improving agent scaffolds.

  • Choose this track if

    Best if you want the highest software/intelligence multiplier across all other frontier domains.

Theory

Week 1: Machine learning as estimation

Reading/source: Deep Learning Ch. 5: learning algorithms, capacity, validation, bias/variance, likelihood

Know: Reconstruct supervised learning as statistical estimation under finite data and distribution assumptions.

Reconstruct: Derive MLE for linear/logistic-style model; bias-variance decomposition intuition; train/validation/test roles.

Do: Train small models on a controlled dataset; deliberately create leakage, overfit, underfit and distribution shift.

Context: Goodfellow-Bengio-Courville remains a clean foundation beneath current foundation-model practice.

Defend: What exactly is learned from data, and what enters through architecture/objective/selection?

Gate: Pass if you can diagnose generalization failure before adding model capacity.

Source: source

Week 2: Deep networks and optimization

Reading/source: Deep Learning Ch. 6-8: feedforward nets, regularization, optimization

Know: Understand computation graphs, backpropagation, normalization/regularization and optimization pathologies.

Reconstruct: Derive backprop for a 2-layer MLP and SGD/Adam-style update logic from chain rule.

Do: Implement a small network from matrix operations; compare optimizer/regularization choices with controlled ablations.

Context: Modern models are larger, but these mechanisms still define the local learning process.

Defend: Why can lower training loss produce a worse research system?

Gate: Pass if ablation conclusions are separated from optimizer noise and seed variance.

Source: source

Week 3: Transformers from first principles

Reading/source: Attention Is All You Need + Stanford CS25 introductory transformer materials

Know: Own self-attention, positional information, residual streams, normalization and transformer block composition.

Reconstruct: Derive scaled dot-product attention dimensions and computational complexity; explain multi-head decomposition.

Do: Implement a minimal transformer and train on a tiny sequence task; instrument attention and residual activations.

Context: CS25 V6 is current in 2026 and surveys frontier transformer work.

Defend: Which transformer components are architectural necessity versus successful convention?

Gate: Pass if implementation matches derivation and you can remove/alter components predictively.

Source: source

Week 4: Scaling, data and evaluation

Reading/source: CS25 scaling/frontier talks + empirical scaling literature; design your own scaling experiment

Know: Reason about compute/data/model tradeoffs and distinguish capability scaling from evaluation artifact.

Reconstruct: Derive log-log power-law fit and uncertainty; calculate compute/memory bottlenecks for attention/training.

Do: Run a tiny scaling study across model/data sizes with held-out tasks and confidence intervals.

Context: Use current CS25 talks as hypotheses, not authority.

Defend: When is extrapolating a scaling curve scientifically unjustified?

Gate: Pass if extrapolation states domain, uncertainty and alternative explanations.

Source: source

Decision systems

Week 5: Reinforcement learning primitives

Reading/source: Sutton & Barto: MDPs, returns, value functions, Bellman equations

Know: Formalize sequential decisions, delayed consequences, state/action/reward and policy evaluation.

Reconstruct: Derive Bellman expectation/optimality equations and tabular policy/value iteration.

Do: Implement dynamic programming and tabular Q-learning on a small environment; compare model-based/model-free.

Context: These equations underlie much later agent/RLHF/post-training work.

Defend: What is hidden when a complex objective is compressed into scalar reward?

Gate: Pass if you identify reward misspecification before tuning the policy.

Source: source

Week 6: Policy gradients and actor-critic

Reading/source: Berkeley CS285 2026: imitation/policy gradients/actor-critic lectures

Know: Understand gradient estimators, variance reduction, advantage functions and on/off-policy tradeoffs.

Reconstruct: Derive REINFORCE estimator and baseline invariance.

Do: Implement a small policy-gradient agent; plot variance, learning curves and failure seeds.

Context: CS285 Spring 2026 includes modern deep-RL and LLM-RL project options.

Defend: What evidence distinguishes policy improvement from lucky trajectories?

Gate: Pass if claims survive multiple seeds and a held-out environment perturbation.

Source: source

Week 7: Offline RL, model-based RL and planning

Reading/source: CS285 2026 modules on model-based/offline RL

Know: Choose between planning, learned dynamics, offline data and direct policy optimization.

Reconstruct: Derive model-prediction-error compounding and one conservative offline-RL intuition.

Do: Train a learned dynamics model; plan through it; expose a model-exploitation failure.

Context: Agent systems increasingly combine learned models with explicit search/planning.

Defend: Why does better one-step prediction not guarantee better planning?

Gate: Pass if you construct and detect a planner exploiting model error.

Source: source

Week 8: LLM post-training as decision optimization

Reading/source: CS285 LLM RL project materials + instruction-following/RLHF papers selected by learner

Know: Connect preference data, reward modeling, policy optimization and evaluation.

Reconstruct: Derive pairwise preference likelihood and KL-regularized policy objective intuition.

Do: On a small open model or toy policy, run preference optimization or simulate reward-model gaming.

Context: Use small-scale reproduction; do not mistake proprietary frontier recipes for first principles.

Defend: What does preference optimization optimize when preferences are inconsistent or strategic?

Gate: Pass if evaluation includes disagreement and reward-hacking tests.

Source: source

Agents

Week 9: Tool use and agent-computer interfaces

Reading/source: SWE-agent paper; inspect ACI design choices

Know: Treat the interface/environment as part of agent intelligence rather than a neutral wrapper.

Reconstruct: Reconstruct agent loop: observe -> reason/plan -> action -> environment -> test/feedback.

Do: Build a sandboxed coding agent for a small repository benchmark; change one interface primitive and measure impact.

Context: SWE-agent showed interface design can materially change coding-agent performance.

Defend: Did the model improve, or did you make the environment easier to act in?

Gate: Pass if interface ablation has reproducible effect and no hidden benchmark leakage.

Source: source

Week 10: Memory, planning and long-horizon tasks

Reading/source: Current agent patterns + SWE-Bench-style long-horizon evaluation

Know: Design memory/state representations and decomposition for tasks longer than a single context.

Reconstruct: Formalize task graph, checkpoint state and termination conditions.

Do: Agent solves a multi-file/multi-stage synthetic task with persistent structured memory; test memory corruption.

Context: Long-horizon coding benchmarks emphasize realistic multi-step work rather than isolated puzzles.

Defend: What should the agent remember, forget, verify, or externalize?

Gate: Pass if memory improves held-out performance without merely storing solutions.

Source: source

Week 11: Multi-agent orchestration

Reading/source: Anthropic 2025 multi-agent research engineering report

Know: Understand orchestrator-worker decomposition, parallelism, context partitioning, coordination cost and economics.

Reconstruct: Derive a simple expected-value model for parallel agents including cost, overlap and failure probability.

Do: Build 3-5-agent research swarm on a benchmark with independent subproblems; compare single-agent quality/cost.

Context: Anthropic reports multi-agent research excels on breadth-first parallelizable tasks but costs many more tokens.

Defend: When does adding agents reduce performance?

Gate: Pass if you identify a task class where multi-agent is worse and explain why.

Source: source

Week 12: Agent evaluation and falsification

Reading/source: SWE-agent + your own adversarial eval suite

Know: Design evaluations that detect shortcuts, contamination, flaky tests, evaluator bias and coordination failure.

Reconstruct: Derive confidence interval for pass rate and a paired-comparison test across agent variants.

Do: Create a hidden-test benchmark with injected misleading cues and flaky tools; evaluate two architectures.

Context: Agent capability claims are only as good as evaluation validity.

Defend: What would make your leaderboard ranking reverse?

Gate: Replication Gate: publish benchmark, seeds, traces, cost and failure taxonomy.

Source: source

Autonomous science

Week 13: Scientific-agent architecture

Reading/source: The AI Scientist (2024): idea -> code -> experiment -> paper -> review

Know: Decompose scientific work into hypothesis generation, experiment, analysis, writing and critique.

Reconstruct: Formalize a closed-loop experiment cycle with state, evidence and stop conditions.

Do: Reproduce a simplified autonomous experiment loop on a toy ML/science problem.

Context: AI Scientist demonstrated end-to-end automation within ML domains.

Defend: Which stage produces evidence versus merely prose about evidence?

Gate: Pass if every claim in generated report links to executable experiment output.

Source: source

Week 14: Agentic tree search for research

Reading/source: AI Scientist-v2 (2025)

Know: Understand branching experiment search, experiment-manager roles and automated review feedback.

Reconstruct: Derive exploration/exploitation tradeoff for experiment tree under a fixed compute budget.

Do: Implement a tiny experiment tree manager; compare greedy vs diverse search.

Context: AI Scientist-v2 reported fully autonomous workshop-paper generation and peer-review acceptance.

Defend: How do you prevent search from optimizing the automated reviewer instead of science?

Gate: Pass if reviewer-gaming test is included and independent metrics disagree constructively.

Source: source

Week 15: Causal and experimental discipline for agents

Reading/source: Reuse Wave-1 causal inference + metrology; autonomous-science experiment design

Know: Require agents to specify estimand/hypothesis, uncertainty, controls and alternative explanations before experiments.

Reconstruct: Regenerate target-trial/DOE templates adapted to computational experiments.

Do: Agent must pre-register experiment table, run controls, and update belief after null/negative results.

Context: This week deliberately adds scientific discipline the agent papers often leave implicit.

Defend: Can your system learn from a negative result without rewriting the goalpost?

Gate: Pass if null results survive into memory/report and alter subsequent search.

Source: source

Week 16: Verification, provenance and research integrity

Reading/source: Tao 2026 'Mathematics in the age of AI' + Leiden Declaration

Know: Separate problem-solving abundance from explanation, taste, agenda setting, verification and community values.

Reconstruct: Write explicit objective decomposition for 'good research': correctness, novelty, explanation, reuse, importance, trust.

Do: Take an AI-generated proof/result and build a provenance/verification dossier plus human-readable explanation.

Context: Tao conditions on research-level AI capability and asks what mathematical research is for; Leiden articulates human-centered principles.

Defend: If proof/solution generation becomes cheap, what remains scarce and worth training?

Gate: Pass if system distinguishes verified result, explanatory value and research priority.

Source: source

Self-improvement

Week 17: Empirical self-modification

Reading/source: Darwin Gödel Machine (2025)

Know: Understand self-modification as search over agent code with held-out empirical evaluation and archive diversity.

Reconstruct: Formalize parent selection -> modification -> validation -> archive loop and selection-bias risks.

Do: Implement a sandboxed toy agent that may change prompts/tools/strategy code but not evaluator; use held-out tests.

Context: DGM reported large coding-benchmark gains through iterative self-modification with safety precautions.

Defend: What exactly improved: agent, benchmark adaptation, evaluator exploit, or compute allocation?

Gate: Pass if candidate modifications are frozen before disjoint held-out evaluation.

Source: source

Week 18: Meta-level self-improvement

Reading/source: Hyperagents (2026)

Know: Distinguish improving task behavior from improving the process that generates future improvements.

Reconstruct: Draw editable task-agent/meta-agent boundary and identify recursive failure channels.

Do: Allow toy meta-agent to modify its own search strategy; compare transfer of improvements across two task families.

Context: Hyperagents extends DGM by making meta-level modification procedures editable.

Defend: How can you tell self-acceleration from ordinary cumulative engineering?

Gate: Pass if improvement transfers to a task not used in selection and costs are normalized.

Source: source

Week 19: Containment and non-regression

Reading/source: DGM safety discussion + sandbox/test engineering

Know: Design capability-growth experiments with immutable evaluation, permissions, rollback and audit trails.

Reconstruct: Derive reliability of a regression suite under correlated/hidden failures.

Do: Build signed immutable evaluator + workspace sandbox + rollback; inject a malicious/self-serving mutation.

Context: Recursive systems magnify evaluator and permission mistakes.

Defend: Which components must remain outside the self-modification boundary?

Gate: Pass if system rejects a benchmark-improving modification that violates a safety invariant.

Source: source

Research apprenticeship

Week 20: Reproduce a frontier agent result

Reading/source: Choose AI Scientist-v2, SWE-agent, multi-agent research, DGM or a recent open agent paper

Know: Learn to reproduce claims before proposing novelty.

Reconstruct: From paper alone, reconstruct hypothesis, experimental design, baselines and key metric.

Do: Reproduce one central claim at smaller scale; document all divergences from original.

Context: Current frontier papers become raw material for verification practice.

Defend: Which result survived reproduction and which depended on undocumented choices?

Gate: Extension Gate: independent reproduction + one falsifying/negative experiment.

Source: source

Week 21: Find a real bottleneck

Reading/source: Failure mining from your reproduction traces

Know: Convert empirical failures into a ranked research agenda rather than feature brainstorming.

Reconstruct: Build causal/fault tree of failures and quantify frequency × severity × tractability.

Do: Create dataset of 50+ failures; cluster by mechanism; select one target with a falsifiable intervention.

Context: Strong research questions often come from stable failure modes, not fashionable architectures.

Defend: Why is this bottleneck upstream rather than a symptom?

Gate: Pass if proposed intervention predicts where it should not help.

Source: source

Week 22: Independent extension

Reading/source: Your chosen primary paper + adjacent literature

Know: Design a minimal novel change that tests a mechanism, not just increases compute.

Reconstruct: Write one-page theory-of-change with expected positive, null and negative outcomes.

Do: Implement extension with frozen baselines and pre-registered metrics.

Context: Novelty is judged against literature and mechanism, not code size.

Defend: What result would make you abandon your favored explanation?

Gate: Pass if change beats or falsifies hypothesis on held-out tasks with cost accounting.

Source: source

Capstone

Week 23: Integrated autonomous research system

Reading/source: All track sources

Know: Integrate model/agent/tooling/memory/evaluation/verification into one bounded system.

Reconstruct: Regenerate complete architecture and trust boundaries from blank page.

Do: Capstone: agent/swarm autonomously investigates a scoped technical question, produces evidence, and survives independent re-run.

Context: The system may automate work; it may not automate away the standard of evidence.

Defend: Where can the system deceive itself without any component 'lying'?

Gate: Systems Gate: external reviewer can reproduce outputs from clean environment.

Source: source

Week 24: Research defense and next 12-month program

Reading/source: Tao + Leiden + all empirical work

Know: Defend what the system can do, cannot do, and what research should be done next.

Reconstruct: Write propositions: capability claims, evidence, assumptions, counterevidence, uncertainty.

Do: Produce technical paper, artifact, benchmark card, safety case, cost ledger, negative-results appendix and 12-month roadmap.

Context: The final product is a research program, not a demo reel.

Defend: If models become 10× more capable next year, which parts of your expertise become more valuable?

Gate: Capstone Gate: oral defense by technical reviewer + red-team reviewer + domain stakeholder.

Source: source

Research gates

G1 Replication

Required performance: Reproduce a central claim from an open agent/autonomous-science paper on disjoint held-out tasks.

Minimum artifacts: Code/env lockfile; benchmark; raw traces; seeds; cost ledger; discrepancy log.

Pass criterion: No novelty credit until a material claim survives reproduction.

G2 Extension

Required performance: Target one stable failure mechanism with a minimal falsifiable intervention.

Minimum artifacts: Failure dataset; preregistered metrics; frozen baselines; negative-result condition.

Pass criterion: Must predict where the intervention will not help.

G3 System Closure

Required performance: Integrate tools, memory, orchestration, evaluation, provenance, sandbox and rollback.

Minimum artifacts: Architecture; trust boundaries; immutable evaluator; independent clean rerun.

Pass criterion: System must reject at least one benchmark-improving unsafe/regressive change.

G4 Research Defense

Required performance: Defend capability, limits, value and next research agenda under faster future models.

Minimum artifacts: Technical paper; eval card; safety case; cost profile; 12-month roadmap; oral defense.

Pass criterion: Must distinguish correctness, explanation, importance and automation value.

Frontier technologies primarily routed here