A source-grounded style guide drawn from ten textbooks or monographs, one set of theory lecture notes, one short algorithm paper, and two technical-writing guides. It distinguishes chapter architecture from the sentence and paragraph craft that makes technical reasoning readable.
The original comparison used 12 supplied PDFs. I inspected each file's front matter or introduction, its contents or section map, representative technical pages, and, where present, examples, exercises, labs, notation guides, or chapter notes. I also inspected rendered pages from six books to check how figures, equations, headings, and running text share the page. For the prose analysis, I returned to at least two passages in every PDF and examined how sentences introduce terms, carry an inference, interpret mathematics, and qualify claims. I subsequently read the two linked technical-writing guides in full. The PDF comparison remains a sample, not a cover-to-cover assessment of the books' technical claims.
Each observation below names a PDF-viewer page. The links target free online copies supplied by authors, publishers, or their institutions.
“Use in the agent notes” is my editorial inference from the sampled style. It is not a claim that the source authors prescribe an agent curriculum.
Borrow teaching methods, not distinctive prose or page design. Prefer a synthesis suited to advanced graduates who have used agents but have not built them.
No single reference has the right mix by itself. The most useful combination is the problem-to-model progression of Shalev-Shwartz and Ben-David, Bishop's and Murphy's interpreted examples, Bach's explicit assumptions and limitations, Mohri's proof structure, and the operational specificity of Mitchell, ISLP, and Phuong–Hutter. The prose should be more expansive than the most compressed theory references.
Several books begin with an intelligible task before introducing notation. The papaya classification problem becomes a formal learning model; a diagnostic test becomes a Bayesian calculation. Agent chapters can begin with a concrete failure or decision—such as a tool timeout after a possible purchase—and then derive the state, observation, authority, and verification machinery needed to reason about it.
A unifying lens lets a reader connect later results to earlier ones. Murphy uses probabilistic modeling; Bach organizes learning theory around estimation, approximation, and optimization; Hardt and Recht move from prediction to action. The agent notes should repeatedly return to a run's state, policy, observations, resources, authority, and outcome. New mechanisms should change a specified part of that model rather than appear as unrelated product features.
A notation guide helps, but the first local use still needs an explanation. State what each variable ranges over, what an operator means, and which assumptions make a claim true. For agent systems, distinguish an actual environment state from an observation, a model proposal from an executed action, and a successful tool response from a verified outcome. An algorithm should state its inputs, outputs, parameters, and permitted effects.
The theory-oriented books name a proposition or guarantee, state conditions, derive it, and explain what the bound means. Murphy's optimization treatment also turns a mathematical monotonicity result into a debugging check. Agent mathematics should do the same: show where a reliability product, a context-cost term, or a permission invariant comes from, then state what the equation predicts and what it does not establish.
Bishop develops several concepts through polynomial curve fitting, and Murphy works Bayes' rule through a diagnostic-test example. ESL compares least squares with nearest neighbors before moving to more complex methods. A useful agent figure should similarly make a state transition, trust boundary, dependency, or cost trade-off visible. Its caption should explain what the reader is meant to infer; it should not merely restate labels.
Mitchell pairs theoretical questions with traces of algorithms. ISLP follows concepts with labs that show code and output. The transformer paper specifies pseudocode with inputs, outputs, parameters, and dimensions. For agent development, include short executable examples of a tool call, context builder, retry state machine, or verifier. Explain how each line implements the preceding model, and keep a consequential operation behind an explicit permission check.
A method comparison is stronger when it explains why a method succeeds or fails in a particular setting. ESL contrasts the structural assumptions of least squares and nearest neighbors; Hardt and Recht distinguish empirical observations from explanations; Murphy's advanced volume ends one causal-estimation section with practical advice. For agents, compare a prompt rule, a tool gate, a sandbox, and a verifier by what each can enforce. Report task success, cost, delay, and risk as separate outcomes.
Formal statements gain teaching value when their idealizations are visible. Bach discusses when asymptotic analysis may mislead at finite sample sizes. Shalev-Shwartz and Ben-David identify the conditions and proof gap for a tighter bound. Telgarsky explicitly marks the weaknesses and omitted topics of his chosen perspective. Agent chapters should distinguish a model property, an engineering check, an observed benchmark result, and an unverified product claim.
Exercises can ask for a derivation, an assumption check, a counterexample, and a practical design decision rather than only arithmetic. ISLP separates conceptual questions from applied work; Mohri connects exercises to chapter notes; Mitchell asks readers to trace an algorithm on altered data. In the main prose, give one claim room to breathe. State a result, explain the mechanism, and then discuss its limitation in separate sentences when those are distinct ideas.
The nine elements above mostly concern the design of a lesson. A second pass through the same twelve PDFs exposed smaller, repeatable moves in their prose. The authors do not share one voice: a proof-oriented text can be terse, an applied text more conversational, and a short algorithm paper unusually compact. The following are editorial inferences from sampled passages, with page-level evidence. They are techniques to adapt, not sentences or mannerisms to copy.
A technical paragraph can state a result, explain the mechanism that produces it, and finish with its consequence or boundary. It need not compress those three jobs into one sentence. Mohri and colleagues move from a bound to its dependence on sample size; Hardt and Recht move from measured accuracy to a narrower conclusion about regularization; Bach follows an asymptotic result with the question of when it applies. For the agent notes, split a claim about a harness from the reason it holds and from the case it leaves unresolved.
A sentence should establish what must be predicted, compared, or proved before the notation becomes dense. The learning questions at the start of Mohri's PAC discussion make the definitions purposeful. ESL asks how a linear model is fitted before giving the optimization rule. Bishop moves from uncertainty in a curve-fitting task to probability theory. In an agent chapter, the reader should know why a belief state or permission predicate is needed before encountering its formula.
Definitions read better when they say what an object does in the problem, not merely the set to which it belongs. Murphy describes a posterior as an updated belief after explaining its normalization. ESL identifies the prediction before discussing whether it is scalar or vector valued. Phuong and Hutter describe the purpose of a positional representation and the consequence of omitting it. In the notes, introduce a proposed action as the model's choice, trusted evidence as the gate's input, and an observation as the environment's report before compressing them into symbols.
Mitchell repeatedly names the hypothesis, hypothesis space, and training example while their relationship is being built. Shalev-Shwartz and Ben-David retain the same terms as they move from a finite class to PAC learnability. Murphy's causal example keeps the education, debt, and wealth variables distinct. Synonym variety is less useful than traceability here. Once the notes define a proposal, do not casually call the same object a command, action, decision, and execution in adjacent sentences; those words may denote different stages of the harness.
Readers need prose where the reasoning changes direction: the assumption that licenses a bound, the inequality that makes an update monotone, or the obstacle that forces a new technique. Murphy explains why the inequalities in an optimization argument hold and turns the result into a code diagnostic. Bach names the difficulty created when the estimator appears on both sides of an inequality. ESL interprets matrix notation geometrically before solving for coefficients. Routine algebra can remain compact; the reason for the consequential step should be explicit.
Immediately after a formula, explain the direction of its dependence and what a reader could verify. Mohri interprets how a generalization bound varies with sample size and hypothesis-class size. Murphy uses a monotonic objective as a debugging criterion. ISLP interprets plotted residuals relative to the fitted function. For a context-cost formula, state which term grows with run length, give a numerical case, and say whether actual billing or quality follows from that calculation.
Bishop uses a simple probability experiment to establish concepts that later apply to less transparent models. Mitchell turns a small hypothesis space into a discussion of search strategies for larger spaces. Murphy's advanced volume uses a numerical causal example to attach meaning to a general structural model. The prose should say which features of the example are essential. For instance, an order cancellation can teach atomic precondition checks; the same pattern applies to any conditional write whose target may change between read and commit.
A transition should show why the next paragraph follows from the last. Mitchell uses the size of the hypothesis space to motivate search. Bishop uses measurement uncertainty to motivate probability. Hardt and Recht introduce a constrained prediction problem before stating the rule that solves it. These are logical bridges, not announcements that a new topic is beginning. A chapter on tool use can move from “the model emits structured arguments” to “valid structure does not establish permission,” which supplies the reason for an execution gate.
Bach states why an asymptotic result may fail to describe a practical sample size. Shalev-Shwartz and Ben-David identify which bound they prove and which tighter proof lies outside the book. Hardt and Recht separate an empirical result from a complete explanation; Murphy's advanced volume says why an estimator with attractive theory may underperform another in practice. An agent note should specify whether a statement is a theorem, a tested behavior of the companion code, an observation in a paper, or an unverified possibility at the point where the statement appears.
The prose surrounding Phuong and Hutter's algorithms names what goes in, what comes out, and how masking changes the result. Mitchell describes how training examples change a learner's hypothesis before following the algorithmic trace. ISLP explains the role of a function argument while teaching a lab. In a graduate agent chapter, explain what a snippet receives, what state it may change, what it returns, and what failure it can expose. Repeating Python syntax line by line would obscure the contract.
The accessible explanations in ISLP and Bishop leave room between a concrete observation and a general statement. Mohri's proofs and Telgarsky's notes show how much can be compressed, but their faster pace is a poor default for readers building an agent harness for the first time. The transferable editorial choice is to vary sentence length, keep one consequential claim per sentence, and use short sentences to land a result. This cadence is a synthesis for these notes, rather than a rule asserted by any one author. Avoid self-referential promises about what the chapter will do when the next object or question can provide the transition.
Compressed: “A status read makes cancellation safe when the request has a revision and an idempotency key.”
Developed: “A status read supplies evidence that the order was unshipped when it was queried. The cancellation service checks the current revision before changing the order. An idempotency key addresses a different failure: a response may be lost after the change commits. Neither mechanism grants the agent permission to cancel. That permission comes from the user's authorization.”
This section analyzes prose mechanisms rather than imitating the source authors' language. Page references identify sampled passages; each recommendation for the agent notes is an editorial inference.
The twelve PDFs establish the chapter's pedagogical depth and logical progression. Two writing guides sharpen the prose at sentence, paragraph, and evidence level. Their advice was read as guidance for scientific writing, then adapted to a graduate textbook whose subject includes probabilistic agents and executable systems.
Heiser's central editorial test is whether the reader has already been given the knowledge needed to understand the next sentence. He also asks for a clear point at sentence, paragraph, and section scale. His advice on concision, active voice, measurements, figures, acronyms, and references provides concrete checks on a draft.
Abbott emphasizes a restrained scientific voice, exact vocabulary, consistent grammar, and figures and equations that function as parts of an argument. His distinctions between proof and empirical support, concrete and vague claims, and stable and historical tense are especially useful when agent-system claims span mathematics, experiments, and product documentation.
| Tension | House rule |
|---|---|
| Brevity versus explanation | Remove words that do no work. Keep the definition, crucial inference, interpretation, and limitation. The reader's request for a little more explanation takes precedence over a short-word-count ideal. |
| Paper structure versus textbook voice | Use informative titles and coherent sections. Do not impose a paper's abstract or section-length rules on a chapter. Avoid “this section will” narration when the subject itself supplies the transition. |
| Absolute word bans versus precise meaning | Abbott discourages words such as “could,” “would,” “should,” and “belief.” Retain a modal when it conveys a defined possibility, counterfactual, or recommendation. Define a belief state as a probability distribution over possible states; do not use a personal belief as empirical evidence. |
| Typographic rules versus useful notation | Prefer an English lead-in and avoid nested parentheses when they improve reading. Keep conventional, locally defined symbols and descriptive code identifiers rather than forcing single-letter names. Mark newly defined terms consistently; the definition matters more than italics or boldface. |
| Different national conventions | Use consistent US English and the Oxford comma for this project. Do not combine incompatible punctuation prescriptions from the two guides. |
The five house rules and conflict decisions above are editorial adaptations, not quotations or claims that both authors endorse every resolution. The two guides are additional references; the twelve PDF portraits below retain their original source-specific scope.
The portraits identify usable habits and a boundary for borrowing them. They describe sampled pages, not every chapter.
Edition / version: Second edition (MIT Press, 2018). Source page
Begins a formal topic with questions, defines a learning model, gives theorem–proof–example sequences, and interprets sample-complexity bounds. Chapter notes and exercises connect the result to literature and extensions. Its deliberately succinct proofs are useful as a structural model; the agent notes should explain more of the intermediate reasoning.
Edition / version: 2006 Springer volume; no numbered edition is stated in this PDF. Source page
Introduces a running example before broadening to probability and decision theory. Figures and captions establish what variables mean before equations become demanding. A dedicated notation section makes conventions explicit. Bishop states that conveying concepts takes priority over maximal mathematical rigor; combine his pacing and visuals with fuller proofs from the theory texts.
Notation p. 11 · Curve fit p. 24 · Probability example p. 32
Edition / version: MIT Press, 2022. This online PDF is dated 18 April 2025; no numbered edition is stated. Source page
Uses a probabilistic lens across topics, restates an equation in words, then tests it on a concrete numerical case. Later derivations connect formal properties to practical diagnostics. Examples, figures, code-generated plots, cross-references, and a notation appendix make a large reference navigable. For agent notes, the most valuable habit is the prose interpretation after each formal result.
Edition / version: MIT Press, 2023. This online PDF is dated 10 December 2025; no numbered edition is stated. Source page
Defines advanced objects systematically, often with a small example before a more general model. Causal graphs and optimization figures accompany derivations; selected sections close with practical advice. Use this as a depth and cross-reference model for topics such as partial observability or causal evaluation, while keeping local definitions in each chapter.
Probability setup p. 39 · Causal example p. 247 · Practical advice p. 1241
Edition / version: Second edition (2009), corrected 12th printing of 13 January 2017. Source page
Defines a prediction problem through scientific examples, then compares methods by the assumptions they impose and the errors they incur. Data examples and exercises make the contrast concrete. Its technical prose can be dense. Borrow the comparative method and empirical figures, but unpack the bridges between assumptions and conclusions for this audience.
Applications p. 20 · Method contrast p. 30 · Exercises p. 58
Edition / version: Python edition; first printing dated 5 July 2023. Source page
Starts with recognizable data sets and visible patterns, explains concepts before equations, then gives labs with executable Python and displayed output. Exercises distinguish conceptual understanding from application. Its accessible cadence is valuable for agent-development labs; the mathematical depth of the agent notes should remain above its introductory level.
Data-first opening p. 12 · Figure interpretation p. 27 · Lab p. 125
Edition / version: Author PDF dated 26 May 2025; no numbered edition is stated. Source page
Starts by asking why the theory matters, provides chapter summaries and prerequisites, and carries assumptions through detailed derivations. It discusses what asymptotic results cannot tell us about finite problems. Adopt its explicit assumption ledger and limiting-case discussion. Add worked agent traces when a proof would otherwise float free of the system being modeled.
Edition / version: McGraw-Hill, 1997; no numbered edition is stated. Source page
Frames machine learning through problems, algorithms, assumptions, and the balance of theory with practice. Definitions and traces are developed on concrete tasks, while exercises vary the task or order of examples. Borrow the progressive algorithm trace. Do not import dated examples or treat historical implementation details as current practice.
Edition / version: 2022 print book; online preprint compiled 8 June 2026. No numbered online edition is stated. Source page
Uses a historical episode to establish why a prediction problem matters, then develops formal criteria and follows their consequences for decisions. Later empirical examples are set beside unresolved theoretical explanations. Borrow the connection from mechanism to social or operational consequence. Use historical framing sparingly when it would delay an agent chapter's main argument.
Historical opening p. 11 · Formal criterion p. 33 · Evidence boundary p. 116
Edition / version: Cambridge, 2014 author PDF; no numbered edition is stated. Its pagination differs from print. Source page
Moves from a concrete classification problem to a formal model, precise quantifiers, theorem statements, and proofs. The reader can see exactly which assumptions a guarantee needs. When the strongest proof needs more machinery, the text identifies the looser result it actually proves. This is the clearest model for formal agent claims, provided every symbol receives local explanation.
Motivating task p. 33 · Definition p. 43 · Proof boundary p. 392
Edition / version: Version 2021-10-27 v0.0-e7150f2d (alpha); lecture notes, not a numbered textbook edition. Source page
Names its chosen theoretical perspective, its weaknesses, and topics it omits. Short propositions and remarks show a route to simplified proofs. The sampled version also contains draft markers and unfinished sections. Use the honesty about scope and the proof simplification; do not imitate the shorthand or unfinished editorial state.
Scope and weaknesses p. 3 · Proposition p. 15 · Draft marker p. 63
Edition / version: arXiv version 1, dated 19 July 2022; a paper, not a textbook edition. Source page
Responds to vague architectural descriptions with formal pseudocode. Algorithms state inputs, outputs, parameters, and dimensions; a notation appendix collects symbols. This is a strong model for a tool or harness contract. As a 16-page paper, it moves too quickly to supply the narrative and exercises needed in graduate lecture notes.
The following sequence preserves the transcript as the topic spine without making the prose a recount of the class. Video timestamps can remain in section headings; the chapter body should read as an independent textbook chapter.
Compressed: “Evaluation should report the vector (S, C, D, R), not collapse everything into a single leaderboard rank.”
Expanded: “Let S measure task success, C resource cost, D delay, and R risk. Report these four quantities separately. A system that improves success by doubling cost may be desirable in one application and unacceptable in another. A single rank hides the information needed to make that decision.”
| Temptation | Editorial decision |
|---|---|
| Copy the shortest proof style | Keep theorem structure and precision, but show intermediate steps and an interpretation. Mohri explicitly favors succinctness; your stated preference favors a little more room. |
| Copy an introductory voice wholesale | Use ISLP's examples and labs while retaining graduate-level models, quantifiers, and failure analysis. |
| Let code replace explanation | Use Phuong–Hutter-style contracts and executable snippets after the model is introduced. Explain what the code guarantees and what it only attempts. |
| Use theory as an unqualified product claim | Mark assumptions and evidence. A theorem about an idealized policy, a benchmark result, and a provider's implementation documentation answer different questions. |
| Imitate distinctive wording or page design | Borrow pedagogical structure, not source prose, diagrams, or typography. The report paraphrases the references and links to freely accessible online editions for inspection. |
The links target freely shared online copies of the supplied editions. PDF-viewer page numbers were checked against the collected files. Google Drive opens ISLP and ESL in its PDF viewer and may not honor a page fragment. The Hebrew University author PDF is listed on the MIT reading list; its host returned a security page during automated verification. This report does not copy or redistribute any PDFs.