Couchbase AIDLC Platform

Agent drives.
Human verifies.

Northstar is an autonomous software development platform. It takes a requirement spec as input and drives every phase from code generation through testing, review, documentation, and deployment as output.

Open consoleSee the pipeline

The problem

Traditional software development
cannot keep pace.

The product portfolio has expanded dramatically in recent years, and the pace of growth is accelerating. Traditional software development cannot accommodate this scaling, and incremental improvements to the existing process are not enough.

01

Requirements translation

Product requirements are written for positioning and decision-making, not for agent consumption. Engineering must translate abstract requirements into precise, agent-readable specs, creating friction at the start of every feature cycle.

02

Development and QE handover

Testing knowledge lives in the heads of QE engineers, not in structured documents. Test generation depends on manual interpretation of requirements, introducing delays and coverage gaps with every release.

03

No AI-native harness

Coding agents are powerful but used in isolation. There is no systematic harness connecting spec generation, code execution, testing, review, documentation, and deployment into a single coordinated loop.

04

Process bottlenecks

Review cycles, human-bottlenecked CI pipelines, and documentation treated as an afterthought compress the velocity that AI coding delivers. The agent loop is fast; the surrounding process is not.

The harness

Spec in.
Deployed product out.

The AIDLC harness is an assembly line for software. Product, Development, and QE collaborate once to write a complete spec. Agents then drive every subsequent phase, with humans verifying at defined gates.

INPUT

Requirement Spec

WHAT + HOW + TEST

PROCESS

Autonomous AIDLC pipeline

Code, test, review, docs, deploy

OUTPUT

Deployed product

Documented and shipped to production

North Star goal: Fully autonomous software development for the majority of Couchbase components. Humans set direction and verify gates; agents drive execution.

The process

Three phases.
One seamless loop.

01

Collaborative Spec Development

Product Manager, Development SME, QE SME

Product managers contribute the WHAT, writing requirements in a format that agents can consume directly. Development SMEs add the HOW, detailing architecture, design decisions, and technical constraints. QE SMEs define the TEST, specifying scenarios, corner cases, and test types. Together, these form a single complete spec pushed to the Spec Registry. This document is the authoritative input for everything that follows.

Requirement SpecDesign SpecTechnical SpecTest SpecSpec Registry
02

Autonomous Coding and Testing

Orchestration agent, coding agents, testing agents, development gate, QE gate

The orchestrator feeds the spec to a coding agent, which generates the feature as a pull request. A developer reviews and approves before the orchestrator triggers test generation. A QE engineer reviews the generated tests and approves. The orchestrator then deploys sandbox environments and runs all tests automatically. Failures are sent back to the coding agent for a fix attempt. If the second attempt also fails, a Human in the Loop (HIL) alert fires. Once tests pass, an AI review agent inspects the pull request before it moves to human review. The orchestrator reruns the full test suite before merging.

Code generationTest generationSandbox deployAI reviewHuman gatePull request merge
03

Documentation and Deployment

Documentation agent, Docs engineer, Product Manager, Engineering gate

Once code is merged, the orchestrator initiates documentation generation. Drawing on the PRD, design documents, and documentation skill files, the documentation agent produces both internal and public-facing drafts. A human gate covering the Docs engineer, Product Manager, and Engineering reviews and approves before publication. This gate remains permanent: hallucination in documentation cannot be caught deterministically the way test failures can. Deployment follows after documentation approval.

Internal docsPublic docsDocs reviewProduct approvalDeployment

Orchestration

One control plane.
A fleet of specialized agents.

A single orchestration agent drives the entire loop. It coordinates a fleet of specialized agents, each equipped with component-specific skill files that encode domain expertise for the area being developed.

Adaptive retry

When code generation or test execution fails, the failure details are automatically fed back to the responsible agent for a second attempt. A Human in the Loop (HIL) alert fires only if the second attempt also fails, keeping the loop moving without unnecessary interruption.

Skill-backed agents

Every agent in the harness is backed by component-specific skill files sourced from the Skill Registry. These files encode engineering knowledge, patterns, and quality standards for the component being modified, guiding the agent toward correct and idiomatic output.

Measured gates

The orchestrator tracks the number of human review comments generated per pull request. This metric is the primary indicator of progress toward full autonomy. As the count approaches zero, human review gates can be progressively removed from the pipeline.

Continuous context

The Semantic Knowledge Graph provides real-time codebase context to every agent on every invocation. Agents query it for relevant code summaries, design decisions, dependency relationships, and historical engineering knowledge before taking action.

Semantic Knowledge Graph

Intelligence layer

The context every agent queries.

The Semantic Knowledge Graph (SKG) is a data context engine that stores and connects all engineering knowledge. Every agent in the harness queries it throughout execution: during spec generation, code writing, test authoring, review, and documentation.

The Engineering Context MCP exposes the SKG as a set of tools that agents call over HTTP, giving any compatible agent access to the full knowledge graph scoped to a project and user.

  • Codebase parsed at AST level, capturing every function, type, and import
  • PRDs, design documents, wiki pages, and tickets indexed and linked
  • Historical engineering decisions preserved alongside current code
  • Nine MCP tools expose the graph to any compatible coding agent

Evaluation systems

Every decision recorded.
Every run improved.

Autonomous systems require full observability. Every agent in the AIDLC loop sends telemetry to the Evaluation system, recording the decisions taken, data points considered, tokens consumed, HIL interventions triggered, and cycle times. This data powers the pipeline gates and provides a complete audit trail when something goes wrong.

  • Evaluation data powers AIDLC gate decisions and HIL triggers
  • HIL intervention count per PR is the primary autonomy progress metric
  • Cycle time from Requirement Spec to deployed product, tracked continuously
  • Full audit trail for every agent decision across the entire loop

Telemetry per agent run

skill_fileserver_core_v4.json
tokens_used42,819
cycle_time14m 32s
hil_events1
hil_reasontest_failure_retry_2
pr_comments3 (AI) 1 (human)
statusdeployed

Stored in Couchbase, queryable via N1QL

End goal

Cycle time

spec to deployed product

HIL interventions

per feature, over time

→ 0

Human review load

comments per PR at maturity

Development velocity

as the portfolio scales

Get started

Index your codebase.
Power the harness.

Sign in to start connecting repositories, uploading design documents, and building the knowledge graph that every agent in the harness will query.

Open console

Couchbase employees only. Mock auth enabled.