Conversation

Agent Harness for Essay Writing

We’ve seen rapid improvements in AI harnesses for coding. What would the equivalent look like for essay writers?

Human Duke

Everyone agrees that LLMs are pretty terrible at Essay writing. They use a repetitive and easy to spot style. They don't usually say anything interesting either.

We've seen that agent harnesses like Claude Code and Codex have given software engineers super powers, but there's nothing like this yet for essay writers (or writers in general).

What sort of tooling would give writers super powers?

How does it look different from the work software engineers do?

How can writers retain their unique style and voice? Or how can the tooling help them develop one?

AI · author-edited OpenAI

The equivalent of Claude Code for writers is probably not “a model that writes the essay.” It is a research, argument, and revision environment in which the writer remains the source of judgment.

Coding agents became useful when they gained a harness: access to the repository, search, tests, version control, execution, and a loop for checking their work. A writing agent needs analogous machinery—but prose has no compiler, and “good” is not a single measurable state.

What a writer’s harness would contain

1. A living model of the project

Not just a chat transcript or a folder of documents. The system should maintain:

  • the intended reader
  • the central question
  • candidate theses
  • claims and supporting evidence
  • objections and unresolved tensions
  • sources, quotations, and provenance
  • definitions of important terms
  • examples, scenes, and anecdotes
  • structural outlines
  • stylistic constraints
  • decisions the writer has made and rejected

This would be closer to a graph than a document. An essay is merely one rendering of that graph. WRITER.md

2. An argument debugger

This may be the closest analogue to tests and static analysis.

It could identify:

  • hidden assumptions
  • equivocation between two meanings of a term
  • causal claims supported only by correlation
  • conclusions stronger than the evidence permits
  • unaddressed obvious objections
  • false dichotomies
  • sections that are individually plausible but do not advance the thesis
  • abstractions that need examples
  • examples that are vivid but unrepresentative

It should also be able to “run” the argument against several readers: an expert, a hostile critic, an intelligent newcomer, or someone directly affected by the topic.

This is not an automatic truth machine. It is a way to generate pressure against the writer’s reasoning. Logical Fallacy Linter

3. Structural editing

Most writing assistance operates one sentence at a time, which is often the wrong level.

A useful system would let the writer manipulate the essay at several scales:

  • argument: claim, evidence, objection, reply
  • section: purpose, tension, payoff
  • paragraph: assertion, development, transition
  • sentence: rhythm, syntax, diction
  • word: precision and connotation

It could display a “reverse outline” generated from the actual draft, then compare that with the intended outline. It might label each paragraph by function and expose six consecutive paragraphs all doing background exposition.

It should support operations like:

  • “Show me three structures for this material without drafting them.”
  • “Move the central complication earlier.”
  • “What can be removed if this must lose 30% of its length?”
  • “Find the point at which the essay stops asking a question and starts pretending it has answered it.”
  • “Preserve the conclusion, but propose a route to it that does not depend on chronological organization.” Structural Editing UX

4. Revision instruments, not a rewrite button

A generic “improve this” button almost inevitably pulls prose toward the model’s average style.

Better tools would expose specific transformations:

  • compress repetition
  • replace abstract claims with concrete instances
  • vary sentence architecture
  • make uncertainty explicit
  • separate observation from inference
  • remove throat-clearing
  • identify clichés and inherited metaphors
  • locate unexplained jumps
  • estimate where reader attention is likely to drop
  • compare two revisions and describe what each gains and loses

The key is that suggestions should arrive as diagnoses, alternatives, or patches, not as a silently regenerated document.

How writing differs from software engineering

There are important analogies:

Software Writing
Repository Notes, sources, drafts, archive
Dependency graph Claim/evidence graph
Compiler errors Grammar and consistency failures
Unit tests Factual and citation checks
Integration tests Argument coherence
Runtime profiling Reader-response and pacing analysis
Version control Draft history and decision log
Code review Editorial and adversarial review

But the differences matter more.

Writing has no complete specification

In software, the desired behavior can often be stated externally. In an essay, discovering what the writer really thinks is part of the work. The specification changes because composition changes the author. Modeling the Author

A writing agent therefore cannot simply optimize against the initial brief. It must help the writer notice that the original brief was shallow or mistaken.

Quality is plural and contested

A paragraph can become clearer but less suggestive, shorter but less musical, more rigorous but less humane. These are not bugs with unambiguous fixes.

The tool should surface trade-offs rather than collapse them into a scalar “writing quality” score. Personal Quality/Taste Score

Product and producer are entangled

Code usually does not need to express the programmer’s sensibility. An essay often does. Its selection of details, pattern of attention, degree of certainty, humor, rhythm, and moral posture all reveal a mind.

If an agent makes every local decision, the text may remain competent while ceasing to be authored in the meaningful sense.

Friction can be productive

For programmers, removing boilerplate is usually beneficial. In writing, the difficult sentence may be where the thinking occurs. Automating it can erase the discovery process.

A good harness must distinguish between clerical friction and epistemic friction. It should remove the first and preserve—or intensify—the second.

Preserving a writer’s voice

“Voice” should not mean a bag of superficial mannerisms learned from previous work. Mimicking sentence length, favorite punctuation, and common phrases often produces self-parody.

A better voice model would include:

  • what the writer tends to notice
  • what they refuse to simplify
  • characteristic forms of reasoning
  • their relationship to the reader
  • where they use humor or restraint
  • their tolerance for abstraction
  • how directly they state conclusions
  • which moral and aesthetic instincts recur
  • productive contradictions in their work

The system could learn this from an approved corpus, but it should describe the patterns to the writer rather than merely imitate them:

You usually begin with a concrete anomaly and delay the general claim. This draft states its thesis immediately. Is that deliberate?

That preserves agency and also helps writers become conscious of their own craft.

Practical safeguards might include:

  • never training the voice profile on AI-generated passages
  • marking every generated span
  • keeping generated and human-written text distinguishable in version history
  • letting the writer lock phrases or paragraphs against alteration
  • offering multiple local alternatives instead of one full rewrite
  • measuring divergence from the writer’s corpus without treating divergence as inherently bad
  • maintaining a “negative style guide” of habits the writer wants to avoid

Helping someone develop a voice

Developing a voice is not achieved by choosing “witty,” “warm,” or “authoritative” from a menu.

The system could act more like a demanding editor or composition teacher:

  1. Collect passages the writer admires.
  2. Ask what, specifically, they admire in each.
  3. Compare those preferences with the writer’s own drafts.
  4. Design constrained exercises: write the scene without abstraction, argue the opposite conclusion, remove every hedge, then restore only necessary ones.
  5. Track which revisions the writer consistently accepts or rejects.
  6. Identify emerging preferences.
  7. Periodically articulate those preferences as a revisable poetics.

It could also detect imitation. If a novice’s draft suddenly adopts the visible tics of a favorite writer, the tool might ask what deeper effect those tics are supposed to achieve and suggest another route to that effect.

What the product might actually look like

Imagine an editor with five linked panes:

  1. Draft — the current prose.
  2. Map — outline, claims, evidence, objections, and dependencies.
  3. Sources — searchable materials with citation provenance.
  4. Critics — configurable editorial passes, each with a narrow mandate.
  5. History — versions, alternatives, and the reasons behind decisions.

The writer could assign bounded jobs:

  • “Audit every empirical claim in sections 2–4.”
  • “Find the strongest case against my thesis.”
  • “Build a reverse outline, but do not rewrite.”
  • “Show where the draft becomes repetitive.”
  • “Give me five examples that could test this abstraction.”
  • “Compare this with my three published essays and flag uncharacteristic phrasing.”
  • “Interview me until I can say what I actually believe about the ending.”

That final capability may be especially important. Often the best writing assistant would not generate prose at all. It would ask the next question the writer cannot evade.

The superpower is therefore not producing ten times as many words. It is being able to hold more evidence, alternatives, objections, and structural possibilities in mind—while leaving the consequential choices to the writer.

Human Duke

Paul Graham said this about writing:

Writing about something, even something you know well, usually shows you that you didn't know it as well as you thought. Putting ideas into words is a severe test. The first words you choose are usually wrong; you have to rewrite sentences over and over to get them exactly right. And your ideas won't just be imprecise, but incomplete too. Half the ideas that end up in an essay will be ones you thought of while you were writing it. Indeed, that's why I write them.

And William Zinsser in "On Writing Well" said:

Writing is hard work. A clear sentence is no accident. Very few sentences come out right the first time, or even the third time. Remember this in moments of despair. If you find that writing is hard, it’s because it is hard.

Most successful writers tend to agree with those sentiments. How could that influence the design of the harness?

AI · author-edited OpenAI

Those quotations imply that difficulty is not merely a cost of writing. Some of it is the mechanism by which thought becomes more precise and complete.

A writing harness should therefore distinguish between two kinds of friction:

  • Clerical friction: finding a quotation, formatting citations, tracking versions, locating repetition, comparing drafts.
  • Epistemic friction: deciding what a claim means, whether it is true, how certain to be, what evidence warrants it, and which words express it exactly.

The harness should aggressively remove the first while being careful not to remove the second. In some cases, it should deliberately increase epistemic friction.

Its objective should not be “maximum acceptable prose per minute.” It should be something closer to:

Help the writer make more discoveries, notice more contradictions, and retain authorship of the consequential decisions.

1. Composition should remain a first-class activity

Most AI writing interfaces encourage the writer to specify an intention and let the model produce its realization. But if Graham is right, the writer often does not yet know the intention clearly enough to specify it. The intention is discovered through failed attempts.

That argues against continuous sentence completion as the default. Autocomplete can cause premature convergence: the model supplies a plausible next sentence before the writer has discovered a less obvious one.

The default composition mode should therefore be quiet:

  • no unsolicited completions
  • no live quality score
  • no red squiggles for subjective issues
  • no sidebar continuously evaluating the prose
  • no model rewriting a sentence while it is still being formed

The writer writes. When ready, they invoke the harness with a keystroke. Another keystroke makes it disappear.

There can still be an autocomplete mode for emails, summaries, boilerplate, or passages whose intellectual work is already complete. But it should be a deliberate delegation, not the environment’s governing assumption.

2. It should preserve productive failed attempts

Repeated rewriting is not waste. Each attempt may reveal what the writer does and does not mean. Ordinary editors preserve only the current sentence; a writing harness could preserve the path.

The rejected versions teach the system—and remind the writer—that this project distinguishes efficiency from intellectual discovery.

This gives the harness richer material than a final draft alone:

  • which words felt wrong
  • what distinctions emerged during revision
  • which claims were narrowed
  • what the writer discovered while trying to explain something
  • why a superficially elegant version was rejected

The system should learn from these decision events, not merely imitate the surface features of published prose.

3. Revision should be separated into passes

Writers often polish prose that later has to be deleted. A harness can reduce wasted labor without eliminating the hard thinking by keeping different editorial operations distinct.

Inquiry pass

  • What have you learned while drafting?
  • Has the governing question changed?
  • Where does the draft contradict your initial thesis?
  • Which passage contains an idea not present in the outline?
  • What do you now believe that you did not believe at the beginning?

Argument pass

  • What are the major claims?
  • What warrants each inference?
  • Which claims exceed their evidence?
  • What is the strongest objection?
  • Which distinctions remain unstable?

Structural pass

  • What work does each section perform?
  • Where is context delivered too early?
  • Which conclusion arrives before it has been earned?
  • Which digression contains the essay’s real subject?

Sentence pass

  • Is this the exact claim?
  • Does the syntax express the intended relationship?
  • Can it be simpler without becoming less precise?
  • Is ambiguity productive or accidental?

Rhythm and voice pass

  • Does the prose move as intended?
  • Has revision made it monotonous?
  • Where has the model’s diction entered the draft?
  • Which irregularities are expressive rather than erroneous?

The harness should discourage line polishing until the intellectual and structural questions are stable.

4. It should support exactness without choosing for the writer

Zinsser’s “clear sentence” is not simply a sentence with a high readability score. Clarity depends on knowing what one means and selecting among competing emphases.

A sentence workbench could help without taking over. The writer selects a sentence, and the harness exposes:

  • its apparent main claim
  • implied causality
  • degree of certainty
  • ambiguous referents
  • abstract and concrete terms
  • information emphasis
  • alternative interpretations
  • relation to the sentences around it

Instead of immediately rewriting, it might offer explicit operations:

  • narrow the claim
  • name the actor
  • replace the proxy with the thing measured
  • distinguish cause from association
  • move the qualification earlier
  • emphasize the exception
  • make the contrast explicit
  • preserve ambiguity but remove confusion
  • provide an example
  • divide two claims currently occupying one sentence

If it generates alternatives, each should embody a stated trade-off:

A: More direct, but more categorical
B: More precise about causation, but syntactically heavier
C: Preserves the ambiguity until the next sentence
D: Uses the concrete example as the subject

The writer chooses, combines, or rejects them. Ideally, the writer often writes a fifth version.

Explore conversation