(translation pending)
When AI agents show up on every SDLC stage, time-to-market drops several fold and quality climbs. But only if humans and agents share the same context, rules, and practices. Not across 7 repos and Notion pages — in a single product.
Each stage's artifact automatically becomes the next stage's input: issue → spec → ADR → PR → image digest → release → SLO → incident → action item → next issue.
Ideas from chats, tickets, and calls are aggregated into a single backlog with semantic embeddings. RFCs are machine-readable.
Spec-driven: spec before code. Epic → story decomposition by an analyst agent. AC in Gherkin or structured YAML.
Architecture-as-code. ADR is mandatory. C4 from text + code analysis. SLOs are defined here, not after prod.
Trunk-based. Golden paths: a new microservice in 30 seconds. Agent runs in firecracker/gVisor with MCP access.
AI reviewer as the first line: style, security smells, missing tests. Human review — for intent and architecture.
Contract tests are mandatory. Self-healing flaky tests. Ephemeral envs per PR. Mutation testing on critical paths.
Hermetic builds, distroless/Wolfi bases, SLSA L3+, Cosign-signed images, SBOM per artifact.
GitOps everywhere. Progressive delivery with auto-analysis against SLO. DB migrations expand/contract. Auto-rollback.
OTel is the only standard. SLO-based alerting. AI agent summarizes alerts (logs + traces + recent deploys).
Blameless postmortems. An incident commander agent spins up a war room, ChatOps commands → actions, action items land in the tracker.
DORA from git/CI/incidents. DevEx surveys. AI clusters feedback into roadmap themes. Every action item has a trace.
No separate AGENTS.md in one place, RFCs in another, runbooks in a third. Service context and rules live in .l1ve/ right next to the code, readable by humans and agents alike.
# .l1ve/context.md — shared context for humans and agents service: payments-api team: billing tier: prod on-call: @alice (primary), @ivan (backup) rules: - PR > 400 LOC requires a split (the splitter agent will suggest one) - migrations via Atlas, separate from deploy (expand/contract) - protected paths: infra/prod/, billing/secrets/ agents: - codegen: sandboxed · MCP[git, fs, lsp, run-tests, docs] - reviewer: read-only · MCP[git, semgrep, sast] - rca: on incident · MCP[loki, prom, tempo, runbooks] slo: - availability: 99.95%/30d → burn alert >2%/h - p99-latency: 250ms → page if >500ms 5min
# .l1ve/agent.yaml — agent-as-catalog-entry
apiVersion: l1vestack.dev/v1
kind: Agent
metadata:
name: codegen-payments
spec:
base: openhands # or aider, cline
model:
provider: h3llo-ai # or anthropic, openai, ollama
model: llama-3.1-70b
context:
rag: # corp knowledge layer
- service:payments-api
- adr:billing/*
- runbook:billing/*
spec_path: docs/specs/
sandbox:
runtime: firecracker
network:
egress: [github.com, registry.h3llo.cloud]
mcp:
- git # scoped: read-only
- filesystem # scoped: ./
- test-runner
- docs-search # ./docs/**
budget:
monthly_tokens: 10_000_000
monthly_rub: 30_000No crippleware. Community — a full SDLC for teams up to 20 people. Enterprise — when you need multi-tenancy and compliance. Cloud — network effects and premium agents that can't be reproduced locally.
Real practices from the h3llo platform team: how we write specs for agents, how AGENTS.md is structured, what we measure once the SDLC is closed.
RFC and PRD templates, YAML frontmatter, machine-readable AC, contract tests as the outcome of a spec.
Domain rules, MCP servers, what NOT to touch, dev commands, test fixtures. With real examples.
Full breakdown of the architectural shift. Which 3 backend plugins we kept, what we rewrote, before/after numbers.
Isolation, network egress policies, scoped tokens, audit log, evals. How we run agents in production.
Lead time, deploy frequency, MTTR, change failure rate. What changes when the SDLC closes.
Ready chart: catalog, scaffolder, agents, MCP gateway, observability. Air-gapped overlay included.
Helm install → ingest git → wire in agents → artifacts are connected automatically. No portal-builder and no 800 lines of YAML.
Download Community →helm install l1vestack h3llo/l1vestack · one command, up in 4 minutes.l1ve catalog ingest --from git@github.com/your/repos · services show up automatically from catalog-info.yaml..l1ve/agent.yaml into the repo · the agent registers itself in the catalog with its runtime and MCP access.An army of AI agents on every stage · DORA through the roof · TTM through the floor. Not across 7 SaaS tools, but in a single product — bring your own model and your own perimeter.