The Problem With Interactive Drafting
Most AI writing tools want you in the seat. You type a prompt, the model writes, you read, you respond, you correct, you wait, you type again. The drafting loop is a conversation that happens in human time, one exchange at a time. For a short memo, that's fine. For a 200-page CSR with thirty-eight sections, hundreds of cross-references, and a bibliography that has to be cleanly numbered at the end, it's a way to spend a week instead of a day.
Regulatory writing has the opposite shape from "AI chat." It's a long, structured pipeline with clear stages: read the protocol, build a retrieval plan, draft each section against module-specific writing rules, cite every claim, resolve cross-references, run a quality pass, package the deliverable. The pipeline doesn't need a human at every stage. It needs a human at specific decision points — and otherwise it should just run.
That's what walk-away drafting is. Initiate the run, close the laptop, return in 1–2 hours to a complete bundle.
What Happens While You're Gone
When you start a walk-away run on a CSR, Asthra moves through the same loop you'd see in an interactive session — but autonomously, in the background, on the server side.
Plan. The agent builds the retrieval plan: which source files map to which sections, which tables anchor which claims, which prior submissions are relevant for narrative tone. The plan is captured as structured state. You can review it after the run, or approve it before walking away.
Retrieve. Semantic retrieval across the locked source set: protocol, SAP, TLFs, prior CSRs, the IB. Tables and figures are parsed alongside the prose. Every retrieval is recorded with passage, page, and offset.
Draft. Section by section. The agent applies the module-specific writing instructions you validated during onboarding — ICH E3 for a CSR, ICH E2C for a PSUR, MDR / MEDDEV for a CER. Every claim is bound to a retrieved passage. Where the source data is missing or contradictory, the agent flags a gap rather than fabricating plausible text.
Cite. Two-level citations are applied — document-level by default, sentence-level on demand. The citation graph is structured state, not strings in the prose.
Hyperlink. Cross-references — every "see Table 14.3.1.2" and "as discussed in §11.4.1" — resolve against stable anchors that were assigned at draft time. The hyperlinks are real Word bookmarks written into the .docx, not plain-text references waiting for a publishing-phase fixup pass.
QC. An end-of-run quality pass runs over the finished draft. It checks cross-reference integrity, statistical-claim source binding, defined-term ordering, citation density, gap detection, and hyperlink resolution. The results land as a structured quality_report.json.
Bundle. The final deliverable is packaged. Four artifacts: the .docx itself, a traceability.json that projects the provenance store, the quality_report.json from the QC pass, and a run_manifest.json capturing the model, code SHA, token usage, and reproducibility metadata.
You return to find all of this ready.
What You Come Back To
The walk-away mode is not the end of the work. It is the start of the writer's review. A regulatory submission is not finished when the agent stops drafting — it is finished when the writer has read every section, applied judgement, accepted or rewritten, and signed off. Walk-away gets you to the starting line for that review, faster than any other path. It does not replace the review itself.
What is different from a traditional first draft is what is waiting for you in the studio when you return. Open the document. The studio loads inside Word and surfaces everything the agent did while you were gone: the full chat history of the run, every retrieval the agent performed, every tool call it made, every gap it flagged, every QC finding from the end-of-run pass. None of it is hidden behind a separate report or a vendor portal. It is all in the same chat surface you would have been looking at if you had stayed in the seat.
That continuity is the point. You did not miss anything. The conversation just continued without you.
From there, the review feels native. You scroll to §9.4, see that the QC pass flagged it for citation density, ask the agent why it cited Protocol §8.2 and not the SAP, get a structured answer, tell the agent to add a cross-reference, watch it apply the change in-document. You summon the FDA Clinical Reviewer persona over §11.2 because the safety narrative feels thin, get back a structured set of findings, accept the ones you agree with. You ask the agent to find recent literature on a specific adverse event for §9.4, approve the lookup, drop the resulting citation in.
The studio is the writer's working environment from this point forward. Walk-away delivers the first draft; the studio is where the draft becomes a submission.
Approval Gates and Resumability
Walk-away doesn't mean the agent has unilateral authority. There are three kinds of pause points built into the loop, and the agent handles them differently depending on whether you're at the keyboard.
Decisions that genuinely need writer judgment — an unusual source-document type the agent isn't sure how to classify, an external literature lookup that needs your sign-off, a section where the available evidence is ambiguous between two readings — surface as approval notifications. The run pauses cleanly, the notification appears in your inbox, and the run resumes the moment you respond.
Decisions inside the agent's scope — typical retrieval choices, ordinary writing decisions, citation placement — proceed without interrupting.
The run is also resumable. If something interrupts the agent halfway through — a tool call fails, a source document was missing a needed table — the run captures its last good checkpoint, raises a flag, and waits. When you come back, you fix the missing input or approve the agent's recovery plan, and the run continues from where it stopped.
What Walk-Away Is Best For
Walk-away drafting earns its keep on three kinds of work.
The first is a fresh first draft of a long document. CSRs and PSURs where you're starting from sources, not editing an existing draft. The pipeline is most valuable when there's a lot of repeated structure to apply.
The second is regeneration after a source update. A new TLF arrives, a SAP amendment lands, an updated safety database export comes in. The agent re-plans against the updated source set, re-drafts affected sections, re-runs the QC pass, and produces a new bundle. Walk away while it works.
The third is overnight batch work. End-of-day, start runs against three CSRs in parallel, walk away. Three bundles waiting in the morning.
It's less of a fit for interactive section-level refinement. If you're editing one paragraph at a time, you want the conversational shape, not the background pipeline.
What the Bundle Looks Like
When the run completes, the bundle that lands has four artifacts side by side. The .docx is the draft itself — hyperlinked, cited, formatted, ready to review. The traceability.json projects the provenance store at run completion — every section maps to its sources, every claim maps to its retrieval, every transaction is timestamped and recoverable. The quality_report.json is the output of the QC pass — pass / flag / gap by check category. The run_manifest.json is reproducibility metadata — model and provider, code SHA, token usage, run duration. If a regulator or an auditor asks "how was this draft produced," the manifest is the answer.
The bundle is portable. It's yours. It survives if the vendor goes away. That's the point.
What Is Next
We are working toward multi-document walk-away — kicking off a CSR + a PSUR + a CER in one batch and getting three bundles back. Cross-document anchors will need to be first-class citizens of the validator. The discipline does not change.
The principle holds: drafting is a pipeline; reviewing is a conversation. The pipeline should run on its own. The conversation should pick up exactly where it left off — every tool call, every retrieval, every flag, all in the studio when you return. The writer's job is to set the inputs, approve the decisions only the writer can make, and apply judgement to the output. Asthra is the environment where that judgement gets applied.
See it on the product page: Walk-away drafting shows the run-state dashboard and the native-continuity handoff. For the broader studio surface, see /product.