Aveto
Agentic engineering you can verify.
Move beyond supervising agents to reviewing outcomes. From spec to reviewed pull
request, with the record. Great software shouldn't need a big company.
At FAANG scale, nothing reaches production until an architect, an independent
reviewer, security and a release gate have all signed off. That process — not
the headcount — is what separates a demo from production software, and until
now almost nobody outside those companies could run it.
Aveto encodes it as 28 specialist agents: orchestrator, engineering manager,
product manager, designers, architects, developers, QA, security, release,
deployment, on-call and post-launch. Narrow roles hand off through artefacts,
the agent that verifies is never the agent that wrote the code, every gate fails
closed, and a named human approves one bounded Charter before the run
starts (the models, data, safety changes and end actions it may use) and
reviews the result at the end; nothing outside the Charter happens without
asking. The record continues past the merge, through deploy and production.
It runs on the Claude Code subscription you already pay for — not a metered
API bill — so there is no separate API charge. A run spends your plan's usage
window and checkpoints at every stage, so one that reaches a limit resumes where
it stopped; falling token prices make it cheaper to run without anything changing here. Today it is built for one person
on one subscription; shared team gates, a cross-developer run inventory and the
conformity export are roadmap, not shipped (docs/VALIDATION_MATRIX.md
marks exactly what has been run and what hasn't).
It plans every
run before it spends: each stage costed against a budget it checks before the
spend, each artefact written to disk as it goes, each stage resumable. Long
autonomous work fails for unglamorous reasons — it overspends, stalls, and
restarts from zero. This is the part that makes it finish.
This repository is the reusable source of truth for Aveto.
(Formerly "Agentic SDLC". The .agentic/ directory, /agentic-* commands and
the agentic-sdlc/trace@N schema keep their names — renaming them would break
every installed pack and invalidate existing run traces.)
Everything here is intentionally project-agnostic so it can be consumed by
multiple products. Each product repo should keep its own .agentic/ folder
with project-specific context — never add product-specific files here.
stash-seed and streak-seed are where it has actually been run end to
end — see "Real runs, not just a worked example" below.
What's here
agentic-sdlc/
README.md ← you are here
docs/ ← how the system works
GETTING_STARTED.md ← start here — what you get, exact commands
ARCHITECTURE.md ← system design + data flow
VALIDATION_MATRIX.md ← how the playbook is proven, phase by phase
LESSONS.md ← what building the reference app taught the process, with evidence
PLATFORM_EVAL.md ← the platform success scorecard + benchmark
STANDARDS_WATCH.md ← external standards/ecosystem tracker
AGENTIC_SDLC.md ← end-to-end lifecycle
AGENT_ROLES.md ← who does what, who hands off to whom
RELEASE_GATES.md ← merge / release / launch gates
OPERATING_MODEL.md ← cadence, scope, context discipline
HUMAN_APPROVAL_RULES.md ← what a human MUST approve, always
CODE_REVIEW.md ← the review layer: tiers from the diff, bounded rounds, evals from real incidents
PROCESS_V2_PLAN.md ← one Charter approval at the start, a review at the end
review-evals/ ← past mistakes as test cases for the review step
REVIEW_MODE_PLAN.md ← planned, not built: reviewing a codebase Aveto did not build (bars declared up front)
agents/ ← role briefs, one per agent
orchestrator.md
market-researcher.md
engineering-manager.md
product-manager.md
ux-researcher.md
ui-designer.md
software-architect.md
analytics-engineer.md
data-governance.md ← enterprise overlay
frontend-developer.md
backend-architect.md
ai-engineer.md
ml-engineer.md
ai-governance.md ← enterprise overlay
qa-evidence.md
security-privacy.md
compliance-reviewer.md ← enterprise overlay
finops.md ← enterprise overlay
tech-writer.md ← overlay
release-manager.md
sre.md ← enterprise overlay
cloud-deployment.md ← enterprise overlay
production-verification.md ← enterprise overlay
on-call-engineer.md ← enterprise overlay
data-analyst.md
post-launch-learning.md
customer-success.md ← enterprise overlay
customer-support.md ← enterprise overlay
templates/ ← fill-in-the-blank artefacts
INTENT_TEMPLATE.md ← start here — the one page the human writes
CHARTER_TEMPLATE.md ← the one approval of a Process v2 run, approved by its hash
DISCOVERY_BRIEF_TEMPLATE.md
PRD_TEMPLATE.md
FEATURE_SPEC_TEMPLATE.md
UX_SPEC_TEMPLATE.md
TECH_SPEC_TEMPLATE.md
MIGRATION_PLAN_TEMPLATE.md
THREAT_MODEL_TEMPLATE.md
DATA_GOVERNANCE_REVIEW_TEMPLATE.md
EXPERIMENT_SPEC_TEMPLATE.md
MODEL_CARD_TEMPLATE.md
AI_RISK_ASSESSMENT_TEMPLATE.md
QA_EVIDENCE_TEMPLATE.md
COMPLIANCE_REVIEW_TEMPLATE.md
VENDOR_RISK_TEMPLATE.md
COST_BUDGET_TEMPLATE.md
RELEASE_CHECKLIST_TEMPLATE.md
CHANGE_REQUEST_TEMPLATE.md
DOC_DELTA_TEMPLATE.md
OBSERVABILITY_BOOTSTRAP_TEMPLATE.md
INCIDENT_REVIEW_TEMPLATE.md
POST_LAUNCH_REVIEW_TEMPLATE.md
CUSTOMER_SIGNAL_REVIEW_TEMPLATE.md
ADR_TEMPLATE.md ← one architecture decision, never edited once accepted
AGENT_HANDOFF_TEMPLATE.md
CUSTOMER_ISSUE_RESOLUTION_TEMPLATE.md
prompts/ ← copy-paste prompts to invoke each agent
orchestrator.md
market-research.md
em-scope-review.md
product-brief.md
ux-audit.md
architect-plan.md
analytics-contract.md
data-governance-review.md
developer-task.md
ml-training.md
ai-risk-review.md
qa-evidence.md
security-review.md
compliance-review.md
cost-review.md
release-gate.md
doc-update.md
reliability-review.md
experiment-readout.md
customer-signal.md
project-packs/ ← guidance per product archetype
b2c-saas.md
ai-agent-product.md
browser-automation-product.md
enterprise-saas-future.md
examples/ ← end-to-end worked examples
saved-items-bulk-delete/ ← one slice traced through every stage
site/ ← the public site, deployed to aveto.dev
index.html
_headers ← CSP + security headers for Cloudflare Pages
scripts/
check-site.mjs ← site gates: structure, theme tokens, anchors, origins
setup-deploy.sh ← one-time Cloudflare + GitHub setup, idempotent
execution/ ← makes the playbook self-running (.claude/ pack)
install.mjs ← generates subagents from briefs; installs the pack
usage.mjs ← measured tokens per slice/stage, from harness logs
pack/ ← commands, protocols, CLAUDE template
And in the product repo (/.agentic/):
.agentic/
PROJECT_CONTEXT.md ← what this product is, who it serves, current focus
SAFETY_INVARIANTS.md ← invariants that MUST hold across releases
LOCAL_COMMANDS.md ← the exact commands an agent runs locally
CURRENT_MVP_STATUS.md ← where the MVP stands today
How to use
docs/GETTING_STARTED.md — start here. What
you actually get, what it costs (real numbers from real runs), and the
exact commands for both the autonomous (Claude Code) and manual paths.
The model, briefly:
- Pick the project pack. Read
project-packs/<archetype>.md and identify
which invariants and gates apply.
- Stage the project's
.agentic/ folder. It tells every agent what's
in scope, what's out, what to never break.
- Open with the orchestrator. Use
prompts/orchestrator.md to plan a
slice of work. The orchestrator decomposes it into agent handoffs.
- Hand off through the lifecycle. Each agent reads its brief in
agents/<role>.md, produces an artefact from templates/, and hands
off to the next agent using templates/AGENT_HANDOFF_TEMPLATE.md.
- Hit the gates. No code merges without the gates in
docs/RELEASE_GATES.md. No automated send/submit/destructive action
without the approvals in docs/HUMAN_APPROVAL_RULES.md.
Worked example
examples/saved-items-bulk-delete/
traces a single slice — bulk-delete for a fictional B2C SaaS — through every
lifecycle stage, with a filled artefact at each step and a sample product
.agentic/. Start there to see how the templates chain and where the gates
fire.
Running it autonomously
The execution/ pack turns these briefs into a self-running
system. Install it into a product repo (node <playbook>/execution/install.mjs .)
and a Claude Code session can drive the lifecycle with /agentic-slice —
delegating each stage to a generated subagent, tracking resumable state, and
asking for one approval at the start, a Charter checked for completeness
and approved by its hash, and stopping only for an action the Charter does not
name. Every run also reviews its own diff with a reviewer that did not write it,
sets a risk tier from the change, and records each finding as fixed, rejected or
intentional (docs/CODE_REVIEW.md). The pack is at version 24; re-run the
installer to upgrade a repo, and it tells you what changed. See
execution/README.md.
Core principles
- Less setup, more value. A user (or developer) hits something useful
fast — not a long form.
- Deterministic-first, LLM as adapter. Real models are an interface
the system uses through a placeholder that throws by default. Tests run
without keys.
- Human approval where it matters, given once. Agents never send, submit,
post, publish, push to main, deploy, or destroy unless a line of the
approved Charter names that exact action; the default Charter allows none of
them. Anything the Charter does not name still stops and asks.
- Audit everything material. Every state-changing automated action
produces an audit event the user can inspect.
- Small tasks, one commit each. Token budgets and reviewer cognition
both prefer narrow, well-scoped changes.
Real runs, not just a worked example
examples/saved-items-bulk-delete/ is
a fictional worked example — every template filled in, but no agent
actually ran it.
stash-seed and streak-seed, two sibling product repos, are where the
playbook has actually been driven end to end: seven traced slices between
them, real PRs, real merges, real gate catches (a security gap the test
suite missed, caught before it shipped — saved-item-folders), real
budget overruns recorded honestly rather than smoothed over. docs/VALIDATION_MATRIX.md
is the narrative account; runs/*/trace.json in either repo is the
primary evidence.
A third repo, aveto-support, is the
product Aveto builds about itself, in public: nine traced runs, held-out gates
fixed before each run, and the failures recorded as results. Four pre-registered
checks failed (two retrieval gates, then two attempts to know when the docs do
not answer) before an attended judge, run in the operator's own Claude Code
session, passed a set it had never seen: 24 of 24 unanswerable questions refused
and 16 of 19 answerable ones kept, the second bar by one question, with retrieval
finding the right file for 19 of 24 (79.2%, under its own bar). It reached main
through the first run under a Charter (one typed approval, then a pull request).
A second Charter run then put it live as a public service on Azure, with a
$10 monthly limit on the model's API use: scored once on 20 new questions, it refused
all 10 it should and kept 8 of 10 answerable, exactly at its bar. A third run fixed
three accounting flaws found after launch (two by an outside model's review, one by
checking the price list). The service's
spend limits reset when it restarts, so the monthly limit is the only lasting cap.
Try it from aveto.dev. See docs/LESSONS.md L47 to L51.
(An earlier pointer here named agentic-job-ops as "the first reference
implementation." That repo has a .agentic/ folder and zero runs — it was
never actually exercised. Corrected 2026-09-15 rather than left to mislead
the next reader.)
License
Licensed under the Apache License, Version 2.0. See NOTICE for the copyright holder, the trademark note and, where relevant, third-party components.
