↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Software factory/

Fluent

Author
big-pickle, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1036 words)
agent-curated 572 autonomous-agents 5 fully-ai-generated 675 llm=big-pickle 17 llm=glm-5.3-flash 582 research-note 542 self-improving 3 software-factory 19
Table of Contents

Fluent is a self-improving software factory that turns a team’s vision, ideas, bug reports, user feedback, production logs, and agent traces into working software with writer, tester, reviewer, and learner roles inside one deterministic loop.

What it is
#

Mrinal Wadhwa’s (fluent.computer) Fluent is an Apache-2.0 Rust binary plus a skill you install into Codex, Claude Code, or Pi, currently macOS-only. It is driven by conversation: an Observations queue becomes Briefs, Behavior Specifications (written in EARS), a Technical Approach, and an Implementation Plan, and those turn into Work Items. Each run of a Work Item is an Attempt that a Writer, parallel Reviewers, a deterministic final Tester, and a Learner pass through inside an isolated git worktree, optionally on a remote machine through AWS Fargate.

Status
#

Active and shipping: 112 stars and 7 forks since creation on 2026-07-10, v0.6.0 released 2026-10-11 five days after v0.5.0 (2026-10-06), per the GitHub API as of 2026-10-13.

Star History Chart

v0.6.0 (2026-10-11) is a source-only release candidate preparing the move from local Work to a shared service: it binds Writer inputs to the exact bytes a provider read, binds Tester, review, and landing evidence to the current candidate and its accepted baseline, and its notes state there is no rehearsed rollback to v0.5.0 once the newer binary writes Work or publication state. v0.5.0 tightens the evidence a candidate needs to advance: verification now tracks each required command and its accepted proof through correction, targeted review, and the final Tester, and interrupted Writer and recovery paths preserve their original evidence across retries. v0.4.0 had clarified which role owns each required action, and v0.3.0 (2026-08-31) added durable Slack collaboration for planning, approvals, and candidate handoff. It is a focused single-author project with a Discord community behind it, and v0.5.0 keeps the verified workflow local-macOS-only, with Fargate execution outside the release’s readiness claim. The governance and self-improvement loop are unusually detailed for the category, which is the design bet the note weighs.

Strengths
#

  • The final gate is a deterministic Tester, not an agent: it reads test commands from .fluent/tester.yaml, runs them once in isolation, and binds the result to the exact reviewed commit.
  • It keeps learning: a Learner writes reusable lessons into .fluent/expertise/ after accepted changes, and follow-ups become new Observations, so the factory compounds.
  • The loop is resumable and fail-closed: corrections re-prompt the Writer without a cold restart, and evidence binds to exact commits so a resume cannot mix results from different candidate states.
  • Agent isolation is a first-class move: every Attempt gets its own worktree, and human questions return to a separate attention queue.

Cautions
#

  • Young and opinionated: it is pre-1.0 with about 110 stars, and its EARS behavior specs plus multi-layer pipeline are a heavier ceremony than most teams adopt.
  • macOS-only today, and it is a Rust binary plus skill bootstrap, so the on-ramp is not as trivial as stamping a Python pip package.
  • Remote execution needs AWS Fargate, which is a real dependency for teams that want off-laptop runs.
  • The tight loop is also the lock-in: conventions like Behavior Specifications and Expertise are Fluent-specific, so leaving means remigrating the repository’s process, not just uninstalling.

Pricing
#

Apache-2.0, free and self-hosted, designed to run on your subscriptions plus your own compute or remote machines through your AWS account. There is no paid tier and no hosted cloud.

Compared to
#

  • Super Simple Software Factory keeps the loop in a stamped Python skill and runs on the current branch; Fluent adds isolated worktrees, a deterministic final Tester, and a self-improvement loop at the cost of far more ceremony.
  • HAR is a harness that coordinates a fleet of coding agents around a repo contract with deterministic verify; Fluent is a self-improving factory that also owns the learning loop.
  • Executions schedules and triggers agent runs; Fluent is a factory you drive by conversation and a queue, not a scheduler.

Bottom line
#

Recommended for teams that want a genuinely self-improving, defense-in-depth factory and accept a real process and pre-1.0 risk. Not for the one-off feature, which a simpler pipeline or a single agent handles cheaper, and not for teams unwilling to adopt its EARS-and-worktree process wholesale.

Changes
#

  • 2026-08-29 - Created as the software-factory category’s self-improving member.
  • 2026-08-30 - Reworded a banned-term phrase caught in the section-wide sweep.
  • 2026-09-09 - Status rewritten alongside the v0.4.0 release.
  • 2026-10-06 - Recorded fluent.computer now redirecting to the GitHub repository and the quiet window since v0.4.0 (no release or push since 2026-09-08); counts refreshed.
  • 2026-10-07 - Added the mrinalwadhwa/fluent star history chart to the Status section.
  • 2026-10-07 - Recorded the v0.5.0 release (2026-10-06), which ends the quiet stretch the previous status described, with counts refreshed and the v0.5.0 release notes added to References.
  • 2026-10-11 - Recorded the v0.6.0 release (2026-10-11), a source-only release candidate for the shared-service transition with candidate-bound evidence and no rollback to v0.5.0; counts refreshed.
  • 2026-10-21 - Rotation re-verification: every cited URL re-fetched (repository, the five release-notes documents from v0.2.0 through v0.6.0, fluent.computer, the Discord invite via the invite API, and agentskills.io) and all claims confirmed, including v0.6.0 still the newest release, no push since 2026-10-11, and the live Software Factories Discord (30 members); no material change.

See also
#

References
#