SYS:NOMINAL
//PROOF //PIPELINE //AGENTS //PLATFORMS //FLEET //FAQ

//AUTONOMOUS SOFTWARE DELIVERY

SOFTWARE THAT
SHIPS ITSELF.

autonom8 orchestrates a fleet of specialized AI agents that spec, build, test and ship your product, end to end, with no human in the loop. Config-driven. Platform-agnostic. Evidence-verified.

//WHAT THE FLEET SHIPS
SPRINT TO RESCUE YOUR WEBSITE

AVG SPRINT: 30 MIN – 3.5 H · EVIDENCE VERIFIED

DEPLOY AUTONOM8 → WATCH THE PIPELINE

EMPTY REPO · SHIPPED PRODUCT · 0 HUMAN INTERVENTIONS · VERIFIED BY BROWSER ORACLE

orchestrator zsh LIVE SESSION

//PROOF

LIVE TELEMETRY

The orchestrator's telemetry wall, the same panels the fleet watches on every run. Figures shown from a demo sprint.

ORCHESTRATOR LOG · OXYGEN/OXYGEN_SITE peak: 44 agents in one sprint
0 TASKS ORCHESTRATED
+0 LINES SHIPPED, REWRITES INCLUDED
first spec parsed PR #8841 opened sprint verified

· pm-agent: decompose A8-142 → 7 tickets, RICE-scored+12−0

· dev-agent: implement checkout flow (flutter)+486−12

· visual-qa-web: 47 checks green, verified+54−34

21 DAYS × 24 HOURS · ~252 SPRINTS
0agent runs
quietpeak
⌘ ORCHESTRATOR · RACE TO GREEN build #53304
0/ 8 platforms green
FAIL FLAKY FIXED PASSING
⌘ CURIOUS-QA · EXPLORATORY SWEEP regressions.txt · 214 fix commits
≈1,204 errors left

//PIPELINE

5 STAGES · 0 APPROVALS

One intent in. One shipped, verified product out. Every stage hands evidence to the next.

01 //PRE-SPRINT

INTENT → PROPOSALS

Proposal processing with PM strategic review and PM↔PO iteration. The loop converges before a single ticket exists.

02 //GROOMING

PROPOSALS → TICKETS

Ticket triage with multi-agent enrichment and PO review. Nothing enters the pipeline unvalidated.

03 //PLANNING

TICKETS → SPRINT

Sprint selection scored with RICE/ICE. The fleet builds the highest-leverage work next, ranked, not guessed.

04 //EXECUTE · P1–P4

SPRINT → CODE

Design strategy locks the UX spec. sprint-architect drafts SPRINT_TODO.json: file ownership, CSS registry, execution order. Decomposition wires dependencies; implement ↔ code review runs to the AC-gate.

05 //VERIFY · P5–P7

CODE → VERIFIED

Playwright integration QA hunts 404s and console errors. Visual QA checks screenshots per platform; design review scores composition. Phase 7 promotes deployed → verified or loops it back.

├── pre-sprint ──▶ grooming ──▶ planning ──▶ execute ──▶ verify ──▶ loop until green

//AGENTS

40 IN THE MANIFEST

Not one generalist. Forty specialists that challenge, build and test each other's work.

//PRODUCT

pm-agent

Scores the backlog with RICE/ICE so the fleet always builds the highest-leverage thing next.

//PRODUCT

po-agent

Turns raw intent into user stories with acceptance criteria sharp enough to test against.

//CORE

sprint-architect

Drafts the master plan: CSS registry, file ownership, execution order. No collisions, no stepped-on toes.

//SECURITY

security-agent

OWASP scans, secret detection, compliance gates. Finds the keys before the internet does.

//QA

visual-qa-web

Drives a real browser: DOM validation, screenshots, console errors. Catches what code review can't see.

//QA

regression-test-agent

Re-runs the whole surface after every change. Nothing ships on a prayer.

//QA

curious-qa-agent

Exploratory testing that hunts edge cases nobody wrote tickets for. Breaks it before users do.

//OPS

devops-agent

Infra, deploys, monitoring, incident RCA. Wakes up so you don't have to.

//MANIFEST

+32 more

bug-miner · doc-fixer · prompt-tuner · metrics-agent · flow-synthesizer · backend-qa →

//PLATFORMS

19 STRATEGIES

One Strategy seam. Point the orchestrator at any stack; it learns the platform from config, not from a rewrite.

WEB FLUTTER IOS ANDROID GO PYTHON RUST JAVA C# PHP SOLIDITY TERRAFORM DATABASE SOLANA AI/ML + YOUR STACK

// add a platform with a config diff, not a fork

//FLEET

5 PROVIDERS · 1 CONTRACT

Claude, Codex, Antigravity, OpenCode and Cursor behind one neutral wrapper contract: github.com/astro44/AI-CLI-Wrappers. Every call returns a JSON envelope with tokens, cost and tool telemetry attached... nothing is spent unseen.

//CLI WRAPPERS · ONE CONTRACT 5 ACTIVE · 1 LEGACY
$ WRAPPERPROVIDERSESSQUOTA
Claude logo$ ./bin/claude.shCLAUDE✓✓
Codex logo$ ./bin/codex.shCODEX✓✓
Cursor logo$ ./bin/cursor.shCURSOR✓✓
OpenCode logo$ ./bin/opencode.shOPENCODE✓✓
Antigravity logo$ ./bin/agravity.shANTIGRAVITY✓✓
$ ./bin/gemini.shGEMINILEGACY

// wrappers are byte-identical to Autonom8-core/bin, verified by check-wrapper-parity --strict

// source → github.com/astro44/AI-CLI-Wrappers
//RESPONSE ENVELOPE · EVERY CALL RETURNS
{
  "response": "...",
  "session_id": "uuid",
  "tokens_used": {
    "input_tokens": 1200,
    "output_tokens": 450,
    "total_tokens": 1650,
    "cost_usd": 0.012
  },
  "metadata": {
    "tool_activity": {
      "call_count": 4,
      "write_count": 1,
      "activity_class": "write_active"
    }
  }
}
⌘ SPRINT COST LEDGER · SAMPLE tokens_used.cost_usd per envelope
$6.90SPEND PER SPRINT · 30 MIN – 3.5 H
11,000,000TOKENS · 42% CACHE-READ
CODEX1,200,000
$2.40
CLAUDE1,000,000
$3.00
CURSOR3,000,000
$0.75
ANTIGRAVITY800,000
$0.40
OPENCODE5,000,000
$0.35

// reconciled per envelope, attributed per token; a two-hour sprint, agents in parallel, costs about a coffee

//QUOTA-STATUS · BEFORE THE BURN
{
  "provider": "claude",
  "quota_exhausted": true,
  "reset_at": "2026-03-12T15:30:00Z",
  "reset_in_seconds": 1800,
  "retry_time": "try again at 3:30 PM",
  "source": "cached"
}

SESSION RESUME

Persona-skip on resume: only new task data enters the prompt. Multi-turn sprints stop paying for their own instructions.

RESPONSE CACHE

Exact + semantic replay of productive calls; the same fingerprint never pays for the same answer twice.

ROLE ROUTING

Per-phase provider pick: design→claude, critic→codex. Scored 50/50 operational health + output quality.

QUOTA GUARDS

--quota-status returns exhaustion + reset time; circuit breakers back off instead of burning through limits.

PROMPT BUDGETS

200k hard cap, warn at 160k, phase-aware section pruning: 14% smaller prompts, truncation 29.5%→18.2%.

LIVE MONITOR

JSONL heartbeats separate "provider is thinking" from "provider is stuck"; healthy reasoning is never killed.

//PRINCIPLES

FOUNDING RULES
01

CONFIG, NOT CODE

Behavior is driven by providers.yaml + project.yaml. The spine has no idea what stack it's building for... and that's the point.

02

AGNOSTIC BY DEFAULT

One Strategy seam, nineteen platforms. Swap the provider or the target with a config diff... never a rewrite.

03

EVIDENCE, NOT ASSERTIONS

Every verdict ships with captured proof: screenshots, logs, DOM state. If it can't be verified, it didn't happen.

04

CLEAN-ROOM PROOF

Autonomy is demonstrated from an empty repository. No operator fixes, no contaminated evidence... the history is the proof.

//FAQ

[ + ] EXPANDS

An assistant waits for a human at every keystroke. An unguarded agent swarm loops until the tokens run out. autonom8 is neither. It runs an opinionated sprint process with clear delineation between tasks, files and roles: scope-enforced, policy-gated, evidence-backed at every stage. Standards like SOC 2 stop being an audit scramble and become a property of the pipeline itself. You review outcomes, not diffs.

A pipeline that decomposes tickets, implements them, tests them in a real browser and gates the deploy... without a person clicking approve between steps. Humans set config and intent; the fleet does the rest.

Nineteen platform strategies out of the box... web, flutter, ios, android, go, python, rust, java, solidity, terraform and more... fronted by Claude, Codex, Antigravity, Cursor and OpenCode CLI wrappers (github.com/astro44/AI-CLI-Wrappers). Everything routes through providers.yaml: adding a platform or a provider is a config change, not a fork.

It doesn't ship. Validation failures auto-create bug tickets with fix hints, the failure routes back to the responsible agent, and self-healing retries run before the next cycle. Failure is a loop, not a wall.

//DEPLOY

STOP MANAGING.
START ORCHESTRATING.

Point autonom8 at a repository. Come back to a shipped product.