Autensa

The World's First Autonomous Product Engine

autensa.com

Your products improve themselves — 24/7 — while you sleep.

Research → Ideation → Swipe → Build → Test → Review → Pull Request — fully automated.

I highly recommend getting Hetzner VPS to run this. You can sign up here.

🎮 Live Demo • Quick Start • Docker • What's New • Features • How It Works • Configuration • Contributors

▶️ Watch the Autensa v2 Introduction

---

🚀 What's New in v2.5.1

Repo Setup & PR Recovery

Autopilot now has a product-level Repo Setup tab that verifies a repository is ready before agents create PR-bound work. It checks authenticated git access, default branch confirmation, GitHub API and PR metadata access, GitHub Actions status, workflow token permissions, PR workflow secrets, and PR workflow variables.

When setup is blocked, users can fix supported GitHub configuration from the UI: set workflow token permissions to read/write, add missing Actions secrets, and add missing Actions variables. Secret values are written directly to GitHub and are not stored in Autensa.

Build Queue tasks with GitHub PRs now include a PR checks recovery panel. Failed checks are classified as retryable, repo setup, or external provider failures, with actions to rerun failed GitHub Actions jobs or rerequest external checks where GitHub supports it.

Private Repo Readiness

Product creation and settings now validate private repos with authenticated git access, detect the remote default branch, and require user confirmation before Autopilot starts. Workspace isolation preflights the selected branch and can use the detected default branch, avoiding main clone failures on repos that use master.

Previous Releases

v2.5.0 — Dispatch & Product Settings Fixes

Per-Task Agent Sessions (#99)

Each dispatched task now gets its own OpenClaw conversation session. Previously, all tasks assigned to the same agent shared one session, causing context to accumulate across tasks until the model's context window was exhausted and the agent stalled. The openclaw_sessions table already had a task_id column — dispatch now uses it for session lookup, session ID generation, and insert. Parallel tasks on the same agent work independently.

Flexible Agent ID Validation (#100)

Agent ID fields now accept both standard UUID format (8-4-4-4-12) and 32-character hex identifiers from the OpenClaw gateway. Previously, Zod's strict .uuid() validation rejected gateway-format agent IDs, causing "Invalid UUID" errors when assigning imported agents to tasks.

Task Delete Button Fix (#111)

The task delete button now shows a loading state ("Deleting..."), disables during the request, and displays inline error messages when deletion fails. Previously, the button had no feedback — if the API request failed or was slow, users saw no response and assumed the button was broken.

Product Pause & Archive (#98)

The Autopilot product settings modal now includes a Status dropdown (Active / Paused) and a Danger Zone section with an Archive button. Paused products stop automated research and ideation cycles. Archived products are hidden from the dashboard but data is preserved. The main product listing now filters out archived products.

v2.4.1 — Community Bug Fixes

• Autopilot model routing — Provider models now route through openclaw/default with the original model in x-openclaw-model, fixing 404 errors on OpenClaw deployments. (@Ahmedkasmi-dev, #109)
• AUTOPILOT_MODEL env var — Removed hardcoded model override in description generation so the shared AUTOPILOT_MODEL config is respected. (@aaronmeza, #116)
• Gateway catalog sync — Local agent role assignments are now preserved during gateway sync instead of being overwritten every 60 seconds. (@cgluttrell, #119)
• Task chat reliability — Agent replies are now captured even without an active SSE connection, and the "waiting" indicator no longer shows stale state. (@heliokeplert-ctrl, #126)

v2.4.0 — Agent Skill Creation Loop

• Agents learn reusable procedures from completed tasks
• Bayesian confidence scoring promotes proven skills
• Matched skills injected at dispatch as primary instructions

v2.0–v2.3 — Full changelog in Releases

• v2.3.x — Idea dedup, operator chat, swipe undo, A/B testing, auto-rollback
• v2.2.x — Preference learning, token tracking, health check endpoints, backup API
• v2.1.x — Server-side pipeline, error reporting, idea badges
• v2.0.x — Session key prefix, dispatch stability, community contributions

v2.0 Highlights

Autensa v2 is a ground-up expansion from task orchestration dashboard to the world's first autonomous product improvement engine. It researches your market, generates feature ideas, lets you decide with a swipe, and builds them — automatically.

🔬 Product Autopilot — The Full Pipeline

The headline feature. Point Autensa at any product (repo + live URL) and it runs a continuous improvement loop:

1. Autonomous Research — AI agents analyze your codebase, scan your live site, and research your market: competitors, user intent, conversion patterns, SEO gaps, technical opportunities. Runs on configurable schedules — daily, weekly, or on-demand.

2. AI-Powered Ideation — Research feeds into ideation agents that generate concrete, scored feature ideas. Each idea includes an impact score, feasibility score, size estimate, technical approach, and a direct link to the research that inspired it.

3. Swipe to Decide — Ideas appear as cards in a Tinder-style interface. Four actions:

4. Pass — Rejected. The preference model learns from it.
5. Maybe — Saved to the Maybe Pool. Resurfaces in 1 week with fresh context.
6. Yes — Task created. Build agent starts coding.

7. Now! — Urgent dispatch. Priority queue, immediate execution.

8. Automated Build → PR — Approved ideas flow through the full agent pipeline: Build agent implements the feature → Test agent runs the suite → Review agent inspects the diff → Pull request created on GitHub with full context.

Your only job is the swipe. Everything else is automated.

📄 Product Program (Karpathy AutoResearch Pattern)

Inspired by Andrej Karpathy's AutoResearch architecture. Each product has a Product Program — a living document that instructs research and ideation agents on what to look for, what matters, and what to ignore. The program evolves as swipe data accumulates: the system learns your taste, not just patterns.

🚛 Convoy Mode — Parallel Multi-Agent Execution

Large features get decomposed into subtasks with a visual dependency graph (DAG). Multiple agents (3–5) work simultaneously with dependency-aware scheduling:

• Parallel subtask execution — Independent pieces run concurrently
• Dependency graph visualization — See what depends on what
• Health monitoring — Detects stalled, stuck, or zombie agents automatically
• Auto-nudge — Reassigns or restarts agents that go dark
• Crash recovery — Checkpoints save agent progress; work resumes from last checkpoint, not from scratch

💬 Operator Chat — Talk to Agents Mid-Build

Don't wait for a PR to give feedback. Two communication modes:

• Queued Notes — Add context ("use the existing auth middleware") that gets delivered at the agent's next checkpoint
• Direct Messages — Delivered immediately to the agent's active session for real-time course correction

Full chat history preserved per task — every message, note, and response.

💰 Cost Tracking & Budget Caps

Granular spend visibility across every dimension:

• Per-task cost tracking — See exactly what each feature costs to build
• Per-product aggregation — Total spend across all tasks for a product
• Daily and monthly caps — Set budget limits that auto-pause dispatch when exceeded
• Cost breakdown API — Detailed reports by agent, model, and time period

🧠 Knowledge Base & Learner Agent

A dedicated Learner agent captures lessons from every build cycle — what worked, what failed, what patterns emerged. Knowledge entries are injected into future dispatches so agents don't repeat mistakes.

📋 Enhanced Planning Phase

Before any build starts, agents run a structured planning phase:

• AI asks clarifying questions about requirements and constraints
• Generates a detailed spec from your answers
• Multi-agent planning specs with sub-agent definitions and execution steps
• Approval gate — you review the plan before any code is written

🔄 Checkpoint & Crash Recovery

Agent progress is saved at configurable checkpoints:

• If a session crashes, work resumes from the last checkpoint — not from scratch
• Checkpoint restore API for manual recovery
• Checkpoint history visible per task

🎯 Preference Learning

Every swipe trains a per-product preference model:

• Category weights (growth, SEO, UX, etc.) adjust based on approvals/rejections
• Complexity preferences calibrate over time
• Tag pattern recognition refines idea generation
• Ideas get sharper with every iteration

🔁 Maybe Pool

Ideas you're not sure about don't disappear:

• Swiped "Maybe" ideas enter a holding pool
• Automatically resurface after a configurable period with new market context
• Batch re-evaluation mode to review accumulated maybes
• Can be promoted to Yes at any time

📡 Live Activity Feed

Real-time SSE stream of everything happening across all products:

• Research progress, ideation cycles, swipe events
• Build progress, test results, review outcomes
• Agent health events, cost updates, PR creation
• Filterable by product, agent, and event type

🛡️ Automation Tiers

Choose your comfort level per product:

Tier	Behavior	Best For
Supervised	PRs created automatically. You review and merge manually.	Production apps
Semi-Auto	PRs auto-merge when CI passes and review agent approves.	Staging & trusted repos
Full Auto	Everything automated end-to-end. Idea → deployed feature.	Side projects & MVPs

🔀 Workspace Isolation

Each build task gets an isolated workspace:

• Git Worktrees for repo-backed projects — isolated branch, no conflicts with other agents
• Task Sandboxes for local/no-repo projects — dedicated directory under .workspaces/task-{id}/
• Port allocation (4200–4299 range) for dev servers — no port conflicts between concurrent builds
• Serialized merge queue — completed tasks merge one at a time with conflict detection
• Product-scoped locking — concurrent completions for the same product queue automatically

📊 Product Scheduling

Configure autonomous cycles per product:

• Research frequency (daily, weekly, custom cron)
• Ideation frequency (after each research cycle, or independent schedule)
• Auto-dispatch rules (immediate on "Yes" swipe, or batch)
• Schedule management UI with enable/disable per schedule

---

✨ Features

Product Autopilot - 🔬 Autonomous market research (competitors, SEO, user intent, technical gaps) - 💡 AI-powered ideation with impact/feasibility scoring - 👆 Swipe interface for instant approve/reject/maybe decisions - 📄 Product Program (Karpathy AutoResearch pattern) - 🎯 Preference learning from swipe history - 🔁 Maybe Pool with auto-resurface - 📊 Configurable research & ideation schedules

Agent Orchestration - 🤖 Multi-agent pipeline (Builder → Tester → Reviewer → Learner) - 🚛 Convoy Mode for parallel multi-agent execution - 💬 Operator Chat (queued notes + direct messages) - 💚 Agent health monitoring with auto-nudge - 🔄 Checkpoint & crash recovery - 🧠 Knowledge base with cross-task learning - 🔀 Workspace isolation (git worktr