👥 Docs team + AI agents 🎯 Continuous docs improvement 📅 2026

Traditional documentation systems require humans to continually identify problems and update guidance. I designed the Documentation Software Factory to automate that feedback cycle using scheduled AI agents, recurring audits, telemetry, and human review to continuously improve documentation quality and discoverability.

Problem

A documentation operating system establishes architecture, workflows, templates, validation, and governance, but none of it stays current on its own. Left unattended, even well-built docs decay:

  • Content drifts as the product moves faster than manual updates
  • Stale links and broken redirects that erode trust
  • Inconsistent terminology creeping back in across contributors
  • Declining discoverability in search and navigation
  • AI answerability gaps, content humans can read but agents can’t reliably use
  • Manual maintenance that scales with headcount instead of the product

The operating system answers “how do we produce good docs?”; the Software Factory answers the harder follow-up: “how do we keep all of it good over time?”

At a glance

Audience

Docs operations team and the agents maintaining the docs

Goal

Automate the feedback cycle so quality improves continuously

Collaboration

Engineering, Growth, DevEx

Tools

Warp (agentic AI), scheduled cloud agents, Git/GitHub, Python, telemetry

Architecture

The Software Factory is built on outer loops: a repeatable pattern where an inner loop does the work and emits signals, those signals accumulate in a structured log, and a scheduled outer-loop agent reads the log to propose improvements to the skills, templates, and guidance that drove the work. Every pass through the loop makes the next one better.

Inputstelemetry, audit logs, PR feedback
→
Agentsscheduled cloud agents
→
EvaluationAFDocs, AEO, SEO, 404
→
Human reviewdraft PRs, approvals
→
Updated skillstemplates and guidance
→
Stronger docs
The outer-loop feedback cycle: signals flow from telemetry and audits through scheduled agents and human review, back into the skills that produce the docs.

Core systems

The really valuable part isn’t any single audit: it’s the architecture that coordinates a set of specialized agents so the whole platform improves over time.

  • Outer loops: The coordinating pattern. Inner loops emit signals; scheduled outer-loop agents read the accumulated signals and propose targeted edits to the underlying skills and templates, so improvements compound instead of resetting each cycle.
  • Scheduled cloud agents: Recurring agents (for example, monthly and quarterly) run unattended, evaluate the docs, and open draft PRs. Now maintenance is a cadence rather than a fire drill.
  • Evaluation audits: A family of specialized checks spanning agent-friendliness (AFDocs), answer-engine optimization (AEO), SEO, cross-link coverage, and 404 detection, each emitting structured signals instead of one-off reports.
  • Signal logs and telemetry: Every run appends structured records to a shared log layer. Accumulated telemetry is what lets agents act on patterns (“this rule is violated repeatedly”) rather than isolated events.
  • Redirect generation: High-confidence fixes, such as redirects for recurring 404s, are drafted automatically from audit data and proposed for review.
  • Skill improvement: The highest-leverage output. The loops fix pages and improve the reusable skills and templates that generate and review every future page.

📸 See the loops, skills, and signal logs live in Warp’s open-source docs repo.

Human in the loop

Automation proposes; humans decide. Every outer-loop agent opens its work as a draft PR that requires human review; nothing auto-merges. Agents handle detection, pattern analysis, and drafting; humans approve the architectural and editorial decisions that define what “good” means. That keeps the factory fast and self-improving without ceding judgment.

Outcomes

  • Continuous improvement replaces periodic cleanup. Docs get better between releases, not just during them.
  • Docs become increasingly agent-friendly as AFDocs and AEO signals feed back into the templates.
  • Maintenance shifts from manual work to recurring evaluation, so effort scales with the product instead of headcount.
  • Reusable skills and templates improve over time, compounding quality across every future contribution.

Lessons learned

  • Durable quality comes from improving the system that produces the work, not from fixing the work piece by piece.
  • Signals only help if they accumulate. Structured logs turn isolated runs into trends agents can act on.
  • Automation earns trust when humans stay the approvers; draft-PR review keeps speed and judgment together.