Skip to content

Latest commit

 

History

History
91 lines (62 loc) · 5.61 KB

File metadata and controls

91 lines (62 loc) · 5.61 KB

Lintel — Platform Vision

Mission

Lintel is the AI-human engineering platform where agents and humans collaborate as teammates within a unified, event-sourced, measurable software delivery lifecycle.

DX, LinearB, and Minware measure engineering performance. Lintel measures and executes — it is both the measurement system and the execution engine. When Lintel detects a quality regression through its metrics projections, it can trigger a remediation workflow automatically.

Core Principles

Event-First

Every state change is an immutable event. Events are past-tense facts — they are never rejected or modified. Commands express intent and may fail. The event store is the single source of truth. All read models (projections), metrics, integrations, and guardrails are derived from the event stream.

Events describe what happened. Guardrails emit commands (never events directly). This prevents circular dependencies.

Command → Handler → (State Change + Event) → EventBus → Subscribers
                                                 ├── Projections
                                                 ├── Guardrail Engine → (New Command)
                                                 ├── Metrics Engine
                                                 └── Integration Sync

Measurable

Every action by every actor (human, agent, system) produces observable data. DORA metrics, agent accuracy, rework rates, cycle times, and cost efficiency are all computed from the event stream without manual instrumentation.

You cannot measure what doesn't exist. The platform follows a strict bottom-up dependency chain:

Layer 6: Continuous Improvement    (learn → feed back → rearchitect)
Layer 5: Measurement & Analytics   (DORA, agent accuracy, team velocity)
Layer 4: Deployment & Observability (deployments, incidents, feature flags)
Layer 3: Guardrails & Safety       (cost limits, escalation, sandbox hardening)
Layer 2: Collaboration & Routing   (teams, channels, integrations, ticketing)
Layer 1: Event Infrastructure      (event bus, subscriptions, projections)
Layer 0: Foundation                (what exists today)

Collaborative

Teams are composed of both humans and AI agents. Agents are first-class team members with the same roles, permissions, and performance records as humans. A team lead can be an AI agent. An agent can approve work items if granted that permission.

Users and agents interact through channels (Slack, Discord, Teams, Web) bound to teams. Projects are ongoing products maintained by teams — they have no end date.

Safe

Guardrails are not blockers — they are enablers. Threshold-based auto-escalation replaces rigid approval chains:

  • Tier 0 (Warn): Notification only — agent rework rate > 20%
  • Tier 1 (Approve): Pause and request human approval — rework rate > 30%
  • Tier 2 (Block): Block and escalate to team lead — rework rate > 50%
  • Tier 3 (Stop): Emergency stop — cost exceeds budget by 200%

Agents operate in sandboxed containers with --cap-drop ALL, seccomp profiles, read-only root filesystems, and no network access (except during controlled clone phase). Cost controls prevent runaway LLM spending. PII is detected and replaced before reaching any LLM.

Differentiation

Dimension DX LinearB Minware Lintel
What it does Measures developer experience Measures engineering metrics Measures time allocation Measures AND executes
AI agents Not tracked Not tracked Not tracked First-class team members with performance records
Event architecture ETL from Git/CI ETL from Git/CI ETL from Git/meetings Native event sourcing — every metric derived from immutable events
Software delivery Measurement only Measurement only Measurement only Full lifecycle: desire → develop → review → deploy → observe → learn
Guardrails None None None Threshold-based auto-escalation with cooldowns
Experimentation None None None Feature flags + A/B testing for agent configurations
Workflow execution None None None LangGraph multi-agent workflows in sandboxed environments
Self-improving No No No Metrics trigger automated remediation workflows

DX/LinearB/Minware are mirrors — they show you what happened. Lintel is a flywheel — it measures, acts, and improves automatically.

Products, Not Projects

Projects in Lintel represent ongoing products. They have no start date, no end date. A product is maintained forever. Work items flow through the product's delivery loop — each work item has a lifecycle, but the product itself is perpetual.

Portfolios group products for cross-product analytics and organizational reporting.

The Software Delivery Flywheel

  +---------+     +---------+     +--------+     +--------+     +---------+     +-------+
  | DESIRE  | --> | DEVELOP | --> | REVIEW | --> | DEPLOY | --> | OBSERVE | --> | LEARN |
  +---------+     +---------+     +--------+     +--------+     +---------+     +-------+
       ^                                                                           |
       +------------- learnings inform next desire --------------------------------+

The delivery loop is fully configurable. Teams define their own phase sequences per product. The default sequence is Desire → Develop → Review → Deploy → Observe → Learn, but phases can be skipped, reordered, or custom phases can be added.

Each phase transition is an event. Each event feeds metrics. Metrics trigger guardrails. Guardrails emit commands. Commands produce new events. The flywheel turns.