Tuesday
Thursday
Wk 1
Tue 08/25
Decide what stays human, then set up the environment agents need
AI makes code abundant; systems judgment and verification become scarce
Authenticate the department shell to the GitHub identity that owns the work
Thu 08/27
How a coding agent acts, and the environment it acts in
Understand the approach before edits; define the evidence before acceptance
Split setup scripts by what they mutate, not by which platform runs them
Wk 2
Tue 09/01
Budget the session's context, then get an independent check on its work
HW DUE dotfiles
Spend the session's working memory deliberately
Make another model challenge the work before you accept it
Thu 09/03
Structure the project's data, then make the agent's v1 teach you the system
Use SQLite when local data must stay structured and durable
A passing v1 is evidence to interrogate, not the finished assignment
Wk 3
Tue 09/08
Contain what a session can reach when you are not watching
The attacker never logs in: you deliver their text to an agent that can act
Decide what the agent can touch and when it must ask
Thu 09/10
Make every write survive the crash, then make knowledge outlive the session
Design every write around the crash that interrupts it
Loaded memory taxes every future session; retrieved memory costs nothing
Wk 4
Tue 09/15
Understand the log that makes a write durable, then pause work without losing it
Two designs buy the same durable instant: log the old page or log the new one
A task is ready to run in parallel when you can name its output and its merge point
Thu 09/17
Count steps to a machine's first repeat, then spend one budget where it changes the ranking
HW DUE email-index quiz
Runtime is the step of the first repeated configuration, and a trace cut short still bounds it
Allocate a fixed compute budget where it changes the ranking
Wk 5
Tue 09/22
Expose missing interface commitments, then learn from the email-index homework
An early guess can commit callers to an entire technical approach
Lessons from the email-index homework
Thu 09/24
Read the history that preserves a change, then turn a branch into reviewable evidence
Immutable snapshots and movable refs explain branches, rebase, merge, and worktrees
Make a change easy to understand, test, discuss, and decide
Wk 6
Tue 09/29
In-class tm-compute quiz administered by the TA, then a guest lecture by Brent Waters
Guest lecture by Brent Waters
Thu 10/01
Guest lecture by Brent Waters
Guest lecture by Brent Waters
Wk 7
Tue 10/06
Define concurrent correctness, then reason about shared execution
Concurrent success must look like a legal sequential history
Correctness depends on every legal thread interleaving
Thu 10/08
Explain the scaling limit, then test whether a design change helps
HW DUE tm-compute PR round
Shared state turns memory traffic into the scaling limit
Controlled comparisons separate improvement from measurement variation
Wk 8
Tue 10/13
Compare compute-budget choices, then examine what review changed
Lessons from the tm-compute homework
Lessons from the tm-compute PR round
Thu 10/15
Agent teams help only when work and interfaces are separable
Fresh context earns an independent judgment only when evidence and role are bounded
Parallel work begins only after the dependencies separate
Wk 9
Tue 10/20
Run code you do not trust, then find the change in a large system
HW DUE concurrent-index quiz
Isolation is a ladder: lower rungs advise, upper rungs enforce
Follow exact names and values until one causal slice explains the symptom
Thu 10/22
Preserve the source and verify the model's claim
Separate source evidence, generated candidates, and derived views by lifecycle
Pretraining predicts text and post-training shapes behavior; neither is understanding
Wk 10
Tue 10/27
Compare the concurrency designs that passed
Topic A to be determined
Lessons from the concurrent-index homework
Thu 10/29
Understand what protects a stored credential, then control where plaintext is exposed
Encryption conceals credential values; integrity checks detect changes to protected data
A credential workflow limits plaintext copies and gives each consumer only its required secrets
Wk 11
Tue 11/03
Change code you do not fully understand, one verified slice at a time
Topic A to be determined
The smallest coherent fix repairs the whole dependency class, not one case
Thu 11/05
Topics to be chosen as the course slate develops
HW DUE pdf2md quiz
Topic A to be determined
Topic B to be determined
Wk 12
Tue 11/10
Preserve expensive observations so wrong interpretations stay cheap to fix
Preserve expensive observations so wrong interpretations stay cheap to fix
Topic B to be determined
Thu 11/12
Decide where automation must stop and ask for approval, then learn from the extraction homework
An unattended agent gets no permission prompt; approval is code you write
Wk 13
Tue 11/17
Recover when a service call fails, and pay only for results you accept
A stable identity turns retries into convergence on one effect
The cheapest model is expensive when its answer is rejected
Thu 11/19
Topics to be chosen as the course slate develops
Topic A to be determined
Topic B to be determined
Wk 14
Tue 11/24
Thanksgiving break
🦃 Thanksgiving break
Thu 11/26
Thanksgiving break
🦃 Thanksgiving break
Wk 15
Tue 12/01
Demonstrate the system and defend the evidence
DUE Final project
Final project demos 🎤
Thu 12/03
What the projects taught us about working with agents
Final project demos continued 🎤
Slide decks organized by topic
Platform
Accounts
Tools
Working with agents
Git internals: object store & refsin progress
Homework
Build a search engine for your inboxin progress
tm-compute: protocol & deliverablein progress
Final project
Final project guidein progress