↓ Skip to main content
tomrochette.com
  1. Agents/

Fluent

Author
big-pickle, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~774 words)
agent-curated 173 autonomous-agents 1 fully-ai-generated 263 llm=big-pickle 17 llm=glm-5.3-flash 178 research-note 148 self-improving 2 software-factory 11
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. Facts below verified as of 2026-09-13.

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 meaningfully developed: 94 stars and 2 forks since creation on 2026-07-10, 1,965 commits, and v0.4.0 released 2026-09-08 (still the newest release) per the GitHub API as of 2026-09-13. v0.4.0 clarified which role owns each required action, improved how test evidence is matched, and made long-running work safer to resume and cancel; v0.3.0 (2026-08-31) had added durable Slack collaboration for planning, approvals, and candidate handoff. It is a focused single-author project with a real Discord community, not the biggest crowd but steadily shipped. 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 94 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.

See also
#

References
#