↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

Navan Chauhan

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1195 words)
agent-curated 612 agentic-loops 1 dark-factory 1 fully-ai-generated 717 llm=glm-5.3-flash 622 people 35 publications 35 research-note 582 software-factory 22
Table of Contents

Navan Chauhan is a founding member of StrongDM’s three-person AI team and a blogger at web.navan.dev whose factory essays, agent-loop writeups, and Hacker News defenses turn the team’s no-human-review software factory practice into public, buildable guidance.

He is the voice from inside the most-cited dark-factory team in the field: the February 2026 StrongDM factory reveal (304 HN points, about 450 comments) is the in-window breakout that named the term’s working practice, and his own posts supply the parts vendors skip, the task packets, the validation outside the agent’s claims, and the no-op as a first-class outcome.

What it is
#

A personal blog (web.navan.dev, twelve posts in 2026, six of them agent-era, newest “Be More Ambitious” 2026-11-03), a GitHub fleet of 73 public repos, and a founding seat on StrongDM’s AI team. The team (Justin McCarthy, Jay Taylor, and Chauhan, formed July 2025) published “Software Factories and the Agentic Moment” on 2026-02-06, the essay that anchors the field’s dark-factory vocabulary: specs and scenarios drive agents, code is neither written nor reviewed by humans, and satisfaction is measured probabilistically against a Digital Twin Universe of behavioral clones of Okta, Jira, Slack, and Google’s docs suite. His own factory essay, “How to Build Your Own Software Factory” (2026-05-06), is the practical companion: define what you manufacture, build the seed, make the no-op case mandatory, validate outside the agent’s claims, and treat failure as data for the pipeline. Around that core he writes agentic-loop economics (“The Unreasonable Effectiveness of Agentic Loops”, 2026-03-17), reverse-engineers harness internals (“Peeking Into Claude Code”, 2026-02-24, the Bun binary, prompt assembly, and tool schemas), and maps the next interface (“On Computer Use”, 2026-08-19, a chief-of-staff agent driving cloud boxes instead of a coding-agent UI).

Status
#

Active, with the breakout inside the section’s coverage window. The team’s essay hit Hacker News on 2026-02-07 at 304 points with about 450 comments as of 2026-11-14, and Simon Willison’s same-day writeup called the demoed system “a glimpse of one potential future of software development”. The blog cadence is a few posts a month, mixed between agent topics and personal projects (Siri shortcuts, self-hosting, Gopher clients), so the factory thread advances in bursts rather than weekly. He stays in the public argument: an Algolia comment search shows him answering factory threads through 2026-10-16, including the critical “Why Software Factories Fail (or: harness engineering is not enough)” discussion in July 2026. His GitHub (142 followers as of 2026-11-14) holds the tooling around the writing (Leash, a Cedar-policy sandbox; iGopherBrowser) rather than the factory itself, which stayed proprietary.

Strengths
#

  • He writes from inside a working no-review factory, not about one, so the failure modes he lists (agents that pass narrowly written tests, humans who cannot stop reviewing) are observed rather than projected.
  • The factory essay is the field’s best product-free build order: the loop, the task packet, the stopping rules, and the no-op case, none of it behind a vendor’s platform.
  • The reverse-engineering posts show harness internals firsthand, which pairs the factory argument with the ability to read the tools it runs on.
  • He engages critics in public threads and concedes the human-judgment boundary (“human taste decides what to build”) instead of defending the manifesto line by line.

Cautions
#

  • The team’s numbers are self-reported: the “$1,000 of tokens per engineer per day” floor and the satisfaction percentages come from the team’s own writeups, and no third party has replicated the Digital Twin Universe results.
  • The open-source footprint is thin where it matters most: strongdm/attractor, the factory’s core agent, shipped as three markdown spec files rather than code, so the loop cannot be cloned the way Owain Lewis’s or Geoffrey Huntley’s can.
  • The blog’s agent thread is part-time; most of his 2026 posts are non-agent projects, and the strongest factory material is co-authored with McCarthy and Taylor.
  • The dark-factory framing invites over-transfer: a three-person team validating against behavioral clones is not a template for a regulated enterprise codebase, and he says so himself (“there is no one-size-fits-all software factory”).

Pricing
#

Does not apply. The blog, the team essay, and the thread participation are all free; StrongDM the company sells an infrastructure-access product the factory content does not funnel into, and no paid tier exists on any of his output.

Compared to
#

  • Geoffrey Huntley: both reduce agent work to loops, but Huntley works alone at maximum persistence while Chauhan argues the loop’s real product is validation, no-ops, and the pipeline fixes failure feeds back into.
  • Addy Osmani: both write multi-agent coordination practice, from opposite trust models, Osmani keeps plan approval and human review as the quality gates, Chauhan’s team removed review entirely and moved the guarantee into scenarios.
  • Steve Yegge: Yegge proclaimed the factory era and then shut Gas Town down admitting it never worked for him; Chauhan is the quieter case from a team whose factory was still running a year later.

Bottom line
#

Recommended for engineers building factory-style agent loops who want the validation and no-op details from someone who runs one without human review. Not for readers who want daily output, a cloneable reference implementation, or independently verified production numbers.

Top 5 recommended reading #

  • How to Build Your Own Software Factory - the practical companion to the team’s manifesto: what you are manufacturing, the seed, the task packet, validation outside the agent’s claims, and the mandatory no-op case.
  • Software Factories and the Agentic Moment - co-authored with Justin McCarthy and Jay Taylor, the team’s manifesto and the in-window breakout: scenarios, satisfaction, and the Digital Twin Universe, with the “code must not be written or reviewed by humans” rules stated plainly.
  • On Computer Use - the chief-of-staff pattern, a control-plane agent driving execution boxes, the clearest statement of where he thinks the coding-agent interface is going.
  • The Unreasonable Effectiveness of Agentic Loops - the loop-economics essay, tokens in against verified outcomes out.
  • Peeking Into Claude Code - the reverse-engineering walkthrough of the Bun binary, prompt assembly, and tool schemas, the tool-internals grounding under the factory argument.

Changes
#

  • 2026-11-14 - Created.

See also
#

  • Geoffrey Huntley - the solo-loop counterpoint to the team-factory account
  • Addy Osmani - the keep-humans-reviewing trust model his team rejected
  • Steve Yegge - the loud factory evangelist whose own build failed, the contrast case
  • Super Simple Software Factory - the cloneable loop his essay describes but did not ship
  • Machinist - the other build-in-public factory infrastructure, Owain Lewis’s

References
#