Example run

The quickest way to understand the system is to watch one full run. This is a real autonomous book the agent produced - The Misprint File - captured in the repo under SampleRun/. The only human input was the seed abstract and the --autonomous flag.

The seed

In a town where every citizen receives a sealed envelope at birth naming the exact date they will die, a young postal carrier finds a misprinted envelope addressed to herself - and sets out to discover who writes the dates, and whether hers can be unwritten.

The kickoff

python writingagent.py new --autonomous --pick 1 --chapters 3 --max-revisions 1 \
  --abstract "<the abstract above>"
python writingagent.py run
python writingagent.py export

That’s it - no further interaction. Here is the actual console log of the run:

== Chapter 1: The Blank Envelope ==
   writing (draft)...
   critiquing...
   verdict=revise confidence=0.90 blocking=2 nits=2
   writing (revision 1)...
   critiquing...
   verdict=revise confidence=0.95 blocking=5 nits=4
   revision cap reached
   autonomous: committing best (unapproved) draft
   [OK] committed chapter 1 (+ summary, canon, index)

== Chapter 2: The Hunter and the Hunted ==
   writing (draft)...
   critiquing...
   verdict=revise confidence=0.95 blocking=4 nits=6
   writing (revision 1)...
   critiquing...
   verdict=approve confidence=0.95 blocking=0 nits=3
   [OK] committed chapter 2 (+ summary, canon, index)

== Chapter 3: The Reprint Protocol ==
   writing (draft)...
   critiquing...
   verdict=revise confidence=0.82 blocking=7 nits=8
   writing (revision 1)...
   critiquing...
   verdict=revise confidence=0.95 blocking=7 nits=6
   revision cap reached
   autonomous: committing best (unapproved) draft
   [OK] committed chapter 3 (+ summary, canon, index)
   [consolidate:final] contradictions=5 unresolved=12
   [production] front=5 back=2 -> manuscript.md
   [learn] +5 skills, 5 watch items; reconciled 5 skills
[OK] Book 'deathdates' complete.

What ran, autonomously

Phase	Outcome
Planner	Proposed 3 directions; auto-picked #1; expanded to a full plan
TOC	3 chapter blueprints (purpose / setup / payoff / dependencies)
Write → Critique	Ch1: revise→revise→cap, committed best. Ch2: revise→approved (0 blocking). Ch3: revise→revise→cap, committed best
Commit	Per chapter: prose + summary + canon (25 characters, world rules, timeline) + index
Consolidation	Found 5 contradictions, 12 unresolved threads (autonomous never pauses - it auto-repairs them)
Production	Generated 5 front-matter + 2 back-matter components; assembled the manuscript
Learner	Distilled 5 reusable craft skills + a 5-item watch-list into the user’s library
Export	manuscript.pdf - 9 pages

What it demonstrates

• True autonomy. The critic is strict - it rejected ch1 and ch3 even at high confidence - and autonomous mode committed the best draft at the cap instead of pausing.
• Self-improving memory. The 5 learned skills are derived from this book’s world (e.g. silent-misprint-handling encodes “the Bureau detains via plainclothes signals, never alarms”) and become retrievable for future thrillers by the same user.
• Production discipline. Facts it can’t know - author, year, ISBN - are left as clear placeholders, never fabricated.
• Continuity tracking. 25 characters, world rules, and a timeline were extracted into queryable canon as chapters committed.

Honest caveats

This sample favored a hands-off demo over polish:

• Consolidation flagged 5 contradictions / 12 unresolved threads - autonomous mode (the default) auto-repairs them. A manual run (run --manual, or autonomous: false) pauses on contradictions (escalate_on_contradiction), and more revision rounds tighten quality.
• The manuscript title appears twice at the very top (assembler H1 + title-page component)
  • a known cosmetic nit.

Browse the complete output - chapters, canon, eval verdicts, learned skills, and the PDF - in SampleRun/.

Prefer a browser? The same run - its live progress, evals, artifacts, cost breakdown, and the Rejected log - is viewable in the web dashboard (writing-agent web). This book run used the defaults, including cost_mode: budget, which routes the judgment nodes to the cheaper flash tier to keep spend down.