Reading in the agentic system

The point: agentic work turns reading into a control skill. You read specifications to preserve intent, agent responses to inspect claims, code and configuration to understand the mechanism, and tests and behavior to establish evidence.

One change, five texts

Code is one representation of a system. The meaning of a change is distributed across artifacts that answer different questions. Reading only code can approve a correct implementation of the wrong request. Reading only a specification can approve a request that does not fit the system. Reading only an agent response can approve a polished but incomplete account.

Specification
Intent and boundary

→

Agent response
Claims and uncertainty

→

Code + config
Mechanism

→

Tests + sensors
Executable evidence

→

Observed behavior
Real outcome

Each arrow is a translation where meaning can be lost. Inspect the important translations; do not reproduce every step manually.

What to read

01 · BEFORE ACTION

Specification

Read for outcome, non-goal, invariant, domain terms, acceptance criteria, and unresolved decisions. Start with what a user or operation should observe—not which files should change.

02 · DURING THE LOOP

Agent response

Read it as a claim map. Ask what it cites, which conclusion is only an assumption, what it omitted, and whether its confidence fits its evidence.

03 · MECHANISM

Code and configuration

Trace input, validation, authorization, transformation, storage, output, and failure handling. Include migrations, flags, permissions, and deployment settings.

04 · PROOF

Tests, sensors, behavior

Read the test's scope and failure condition. Then inspect the boundary that matters: UI, API, permission denial, migrated record, or operational signal.

Read by decision

Decision	Read first	Require from the agent	Not proof
Right problem?	Specification and user journey	Scope, assumptions, open questions	Fluent summary
Plan preserves intent?	Agent plan and cited context	Alternatives and source citations	Confident plan
Safe location?	Entry point, path, boundary, tests	Affected files and unchanged contracts	File list
It works?	Acceptance and failure case	Commands, results, known gaps	“Tests pass”
It remains clear?	Diff and local conventions	Line-specific explanation	Polished prose

Seven skills

Purpose reading

Name the decision an artifact must support before opening it. Stop when the decision is supported, not when the page ends.

Claim tracing

Follow a statement back to a specification, source, code path, command, or run result. An approval should have visible evidence behind it.

Boundary detection

Notice invariants, permissions, ownership, compatibility, and failure paths. Ask what must not change before reviewing what changed.

Questioning

Turn ambiguity into a decision with an owner. “Should failed imports retain partial records?” is better than a longer vague prompt.

Synthesis

Connect the specification, response, code, test, and behavior into one causal story. This is independent understanding.

Calibration

Match confidence to evidence and risk. Say “not established” when the available proof is not enough.

Vocabulary building

Learn the domain and technical terms that carry real meaning. Repeated use matters more than a long word list.

Practical inference

A passing test suite proves only the behavior it exercises. Treat it as a sensor with a coverage boundary, not a universal guarantee.

The four-step reading loop

MapOutcome, path, boundary, evidence.

PredictExpected change, protected invariant, surprise.

VerifyDiff, focused test, one sensor, boundary behavior.

RetainTurn a repeated surprise into a guide or check.

First, draw the smallest model that makes work legible. Next, predict what should change and what must not. Then read evidence in increasing cost order. Finally, classify a repeated surprise: missing intent, rule, retrieval path, test, or sensor. Improve the harness instead of merely reading every future patch more slowly.

Exercises that build the skill

Specification paraphrase: before work starts, write outcome, non-goal, invariant, and deciding question in four lines.
Response audit: mark each sentence in an agent conclusion as supported fact, inference, or unresolved claim. Open one source for every high-impact claim.
Translation check: trace one acceptance criterion through response, code/configuration, test, and observed behavior.
Teach-back: explain the causal chain in four sentences without copying the agent's wording. If you cannot, read the sources again.

Technical English for non-native readers

Reading English may use more working memory, especially when a document mixes domain language, idioms, dense sentences, and unfamiliar technical terms. Do not translate every word. That breaks the causal flow. Use two passes.

Pass one

Get structure and decision

Read headings, first and last sentences, tables, identifiers, examples, and emphasis. State the purpose, two essential terms, and one remaining question in simple English or your strongest language.

Pass two

Resolve only high-value gaps

Mark a word or sentence only if it blocks outcome, constraint, failure condition, instruction, or evidence claim. Guess from context, inspect word parts, check a reliable source, make a task-specific example, then re-read.

Build a working vocabulary

Keep only terms that recur in your real specifications, reviews, and agent responses. Meet each term in reading, writing, speaking, and review. Do not build a word cemetery.

Termidempotent

Plain meaningRepeating the same request has the same effect as doing it once.

Current exampleRetrying this payment callback must not create a second invoice.

ContrastNot the same as “safe to retry” unless errors and side effects are handled.

Your sentence“The import endpoint needs an idempotency key.”

Use the agent as a reading coach

“Explain this in plain English. Keep the technical terms that matter and define them once.”
“Show subject, action, condition, and consequence in this sentence.”
“List five essential terms from this specification with an example from this task.”
“Translate this into Thai and back into simple English. Flag meaning that does not transfer cleanly.”
“Ask three questions that test the decision in this document, not isolated vocabulary.”

Check important translations against the original. Negation, modal verbs such as must and may, quantities, conditions, and exceptions often carry more operational meaning than uncommon nouns.

A 20-minute weekday routine

Minutes	Activity	Result
3	Skim one real artifact and state its decision in simple English.	One-sentence purpose
5	Read one short section closely; mark only language that blocks meaning.	Two to five high-value gaps
4	Use context, word parts, and a reliable reference to resolve gaps.	Plain meanings and an example
4	Ask the agent three comprehension questions; verify each against the text.	Corrected understanding
4	Write or say a four-sentence teach-back.	Outcome, condition, evidence, open question

Use easier material when the aim is fluency. Use harder material when the aim is to learn one important concept. Keep the material connected to current work so every new phrase has an immediate use.

Next experiment

Rewrite the next multi-file task as outcome, non-goal, invariant, and deciding question.
Require the agent to separate facts, inference, and open questions.
Write a prediction before opening the final diff.
Trace one acceptance criterion from specification to observed behavior.
When the same surprise occurs twice, add a guide, test, or retrieval step. When the same English term blocks comprehension twice, add it to the working vocabulary.

Sources

Self-contained reader. No private-vault links, remote fonts, scripts, stylesheets, or images are used.