Skip to main content

comprehensive research guide

GitLab Ultimate Agent Platform - Comprehensive Research Guide

Research Completed: January 7, 2026 Total Research: 50,000+ words across 8 specialized agents Coverage: GitLab Duo, OSSA, MCP, A2A, Drupal, Commercial Frameworks, CI/CD, Cursor


Table of Contents

  1. Foundational Agents & Flows
  2. Agent Interoperability Standards
  3. Drupal Agent Marketplace
  4. Commercial Agent Frameworks
  5. CI/CD Agent Patterns
  6. Platform Agents Repository
  7. Cursor IDE Integration
  8. Implementation Roadmap

1. Foundational Agents & Flows

GitLab Duo Agent Platform Overview

Status: GA (General Availability) January 2026 with GitLab 18.8 Requirements: Premium/Ultimate + GitLab Duo Core/Pro/Enterprise add-on Model: Anthropic Claude Sonnet 4.5

Three Foundational Agents (Beta - Enabled by Default)

1.1 Planner Agent (GitLab 18.6+)

Purpose: Product management and planning workflows

Capabilities:

  • Strategic Planning: Prioritization (RICE, MoSCoW, WSJF), work decomposition, dependency analysis
  • Content Creation: Memos, summaries, issues, epics, features, requirements
  • Optimization: Backlog refinement, effort estimation, scope reduction, MVP definition

Best Practices:

  • Provide context via URLs or filter criteria
  • Specify prioritization frameworks
  • Clarify workflow assumptions

Access: GitLab Duo sidebar or @Planner mentions

1.2 Security Analyst Agent (GitLab 18.6+, Ultimate Only)

Purpose: Security analysis and vulnerability management

Capabilities:

  • Vulnerability assessment and prioritization
  • Risk evaluation (severity, exploitability, business impact)
  • False positive filtering
  • Compliance support (regulatory requirements)
  • Security reporting and remediation planning

GitLab Integration:

  • List all project vulnerabilities
  • Provide CVE data and EPSS scores
  • Confirm, dismiss, or update severity
  • Create/link vulnerability issues

Prerequisites:

  • Security scanning enabled
  • Ultimate tier with Duo add-on

1.3 Data Analyst Agent (GitLab 18.6+, Beta in 18.7)

Purpose: Platform data analysis and visualization

Capabilities:

  • Query data using GitLab Query Language (GLQL)
  • Quantify work items across timeframes
  • Evaluate team productivity
  • Identify workflow patterns
  • Generate embeddable GLQL queries

Example Queries:

How many merge requests were merged this month?
What has @username worked on this month?
Show me open issues with ~priority::1 and ~bug labels
What's the trend of bug creation this quarter?

Limitations:

  • Light aggregation only (100 item limit)
  • GLQL supports specific data areas
  • Cannot output directly to dashboards

Four Foundational Flows

1.4 Software Development Flow (GitLab 18.2+)

Purpose: AI-generated solutions across SDLC

IDE Support: VS Code, Visual Studio, JetBrains Languages: CSS, Go, HTML, Java, JavaScript, Markdown, Python, Ruby, TypeScript

How It Works:

  1. Creates and works through a plan
  2. Stages proposed changes
  3. Maintains contextual awareness
  4. Allows supplementary context from issues/MRs
  5. User retains control (accept/modify/reject)

Security: Time-limited OAuth token, write operations based on user permissions

1.5 Developer Flow (Issue to MR) (GitLab 18.6+)

Purpose: Automatically convert issues into merge requests

Process:

  1. Analyzes issue description
  2. Creates development plan
  3. Generates code structure
  4. Opens draft MR linked to issue
  5. Executes automated pipelines

Prerequisites:

  • Developer role or higher
  • Existing issue with clear requirements
  • Service account access

Best Practices:

  • Keep issues well-scoped
  • Specify exact file paths
  • Include acceptance criteria
  • Provide code examples

1.6 Code Review Flow (GitLab 18.7+)

Purpose: Streamline code reviews with agentic AI

Capabilities:

  • Examines code modifications, discussions, linked issues
  • Enhanced contextual understanding
  • Delivers detailed, actionable feedback
  • Focus on specific aspects (security, performance, maintainability)

Activation:

  • Assign review to @GitLabDuo
  • Mention @GitLabDuo in MR comments
  • Enable automatic reviews at project/group level

Customization:

  • Repository-specific review instructions
  • Focus areas configuration
  • Coding standards enforcement
  • File pattern targeting

1.7 Convert to GitLab CI/CD Flow

Purpose: Migrate Jenkins pipelines to GitLab CI/CD

Process:

  • Converts Jenkins pipeline syntax to GitLab CI/CD YAML
  • Creates MR with converted configuration

2. Agent Interoperability Standards

2.1 OSSA (Open Standard Agents)

Project: https://openstandardagents.org Maintainer: Bluefly.io Status: v0.3.2 stable

Purpose: Vendor-neutral agent specification format - "OpenAPI for AI Agents"

Key Features:

  • 20+ LLM provider support
  • Enterprise security (OIDC, secrets management)
  • Native OpenTelemetry integration
  • Bi-directional conversion with GitLab Duo

Manifest Structure:

name: code-reviewer version: 1.0.0 domain: gitlab subdomain: merge-requests type: worker runtime: typescript access: tier: tier_2_write_standard permissions: [read_code, write_comment] capabilities: - name: code_review version: 1.0.0 skills: [security-analysis, performance-review] llm: model: claude-sonnet-4-5 temperature: 0.3 max_tokens: 8192 observability: tracing: enabled metrics: prometheus logging: json

2.2 MCP (Model Context Protocol)

Specification: https://modelcontextprotocol.io Maintainer: Agentic AI Foundation (Linux Foundation) Status: v2025-06-18 specification

Purpose: Connect AI assistants to tools and data sources

GitLab MCP Integration:

As MCP Server (Exposes GitLab Data):

  • Projects, groups, issues, merge requests
  • CI/CD pipelines, job logs
  • Repository files, commit history
  • Available tools: create_issue, create_mr, trigger_pipeline

As MCP Client (Connects to External Servers):

  • Access Jira, ServiceNow, ZenDesk
  • Enable agents to use external tools
  • Maintain single interface

Configuration Example:

{ "mcpServers": { "gitlab": { "type": "http", "url": "https://gitlab.com/api/v4/mcp", "auth": { "type": "oauth2" } } } }

2.3 A2A (Agent-to-Agent Protocol)

Specification: https://a2a-protocol.org Maintainer: Linux Foundation Status: Backed by 150+ organizations (AWS, Google, Salesforce, ServiceNow)

Purpose: Secure inter-agent communication

Discovery Mechanisms:

  1. Open Discovery (.well-known): Agent Card at /.well-known/agent-card.json
  2. Curated Registry: Search by capabilities, metadata-based discovery
  3. OpenAPI Integration: Each agent's spec for invocation

GitLab Integration Pattern:

Cursor Agent  (MCP)  GitLab Duo Agent Platform
  
GitLab Duo Agent  (A2A)  External Security Scanner
  
Results  (A2A)  Cursor Agent  Display to user

2.4 kagent (Kubernetes-Native Agents)

Project: https://kagent.dev Status: CNCF Sandbox Project Purpose: Kubernetes-native agent deployment

Features:

  • Multi-protocol support (MCP, A2A, ADK)
  • Agent mesh architectures
  • Built on Kubernetes CustomResourceDefinitions
  • Declarative agent deployment

Integration with GitLab:

  • Deploy GitLab Duo agents as Kubernetes resources
  • Use GitLab Agent for Kubernetes (GitOps)
  • Manage agent lifecycle via kubectl

3. Drupal Agent Marketplace

3.1 Drupal 11 AI Ecosystem

Core Modules:

  • AI Module: 48+ AI provider integrations, 4,082 sites using
  • AI Agents Module: Text-to-action agents with tools-calling
  • MCP Server: OAuth 2.1, official PHP SDK
  • MCP Client: Connect to external MCP servers

3.2 Built-in Agents

Three Default Agents:

  1. Field Type Agent: Creates/edits fields on entities
  2. Content Type Agent: Handles node type operations
  3. Taxonomy Agent: Works with vocabularies and terms

3.3 Vector Database Integration

Supported:

  • Milvus (official module): Cosine similarity, Euclidean distance, inner product
  • Qdrant (under consideration): Docker deployment, LangChain/LlamaIndex support
  • pgvector (PostgreSQL extension): Native vector storage

3.4 Agent Marketplace Architecture

Content Type: ai_agent

Fields:

  • Agent Name, Description
  • Agent Type (taxonomy: code_review, documentation, testing, deployment)
  • Capabilities (multi-select taxonomy)
  • Endpoint URL
  • Authentication Method (OAuth2, API Key, OIDC)
  • OpenAPI Spec (file or URL)
  • Version, Provider, Status
  • Supported Languages
  • Cost Per Request, Avg Response Time
  • Usage Count, Rating

JSON:API Discovery:

# Query agents by capability GET /jsonapi/node/ai_agent?filter[capabilities.name]=code_review&filter[status.name]=active # Response includes endpoint URL, auth method, OpenAPI spec URL

MCP Server Integration:

{ "tools": [ { "name": "search_agents", "description": "Search for agents by capability, provider, language" }, { "name": "invoke_agent", "description": "Invoke a specific agent with payload" } ] }

3.5 ECA Workflows (Event-Condition-Action)

Purpose: Visual workflow automation without code

Scale: ~500 actions, ~70 conditions, ~200 events

Agent Chain Example:

Event: New content created
  
Condition: Content type = "Product"
  
Action: Invoke AI Agent to generate description
  
Action: Invoke AI Agent to suggest taxonomy terms
  
Action: Invoke AI Agent to generate alt text for images

4. Commercial Agent Frameworks

4.1 Claude Code for GitLab CI/CD

Status: Beta, maintained by GitLab Purpose: Event-driven AI automation through MR workflows

Setup Options:

Option 1: Claude API (Simple)

claude: stage: ai image: node:24-alpine3.21 variables: ANTHROPIC_API_KEY: $ANTHROPIC_API_KEY script: - npm install -g @anthropic-ai/claude-code - claude -p "Review this MR and implement changes"

Option 2: AWS Bedrock (OIDC)

claude-bedrock: id_tokens: AWS_OIDC_TOKEN: aud: https://sts.amazonaws.com script: - aws sts assume-role-with-web-identity - claude -p "Implement feature"

Option 3: Google Vertex AI (Workload Identity Federation)

claude-vertex: before_script: - gcloud auth login --cred-file=<(cat <<EOF...) script: - CLOUD_ML_REGION=us-east5 claude -p "Review code"

4.2 OpenAI Agents SDK

Languages: TypeScript/JavaScript, Python Purpose: Lightweight, production-ready multi-agent workflows

Core Primitives:

  • Agents: LLMs with instructions and tools
  • Handoffs: Delegation between agents
  • Guardrails: Input/output validation (Zod)
  • Sessions: Automatic conversation history

Basic Pattern:

import { Agent, run } from '@openai/agents'; const agent = new Agent({ name: 'CodeReviewer', instructions: 'Review code for security issues', model: 'gpt-4o', tools: [...] }); const result = await run(agent, userInput);

4.3 Observability & Monitoring

LangFuse (Open Source):

  • OpenTelemetry backend support
  • 100+ LLM provider support
  • Cost tracking with Daily Metrics API
  • Self-hosted or cloud options

Helicone (Open Source):

  • AI gateway with routing, failover, caching
  • One-line integration via proxy URL
  • Semantic caching (20-30% cost reduction)
  • 100+ LLM providers

OpenTelemetry Direct:

import { NodeSDK } from '@opentelemetry/sdk-node'; import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http'; const sdk = new NodeSDK({ traceExporter: new OTLPTraceExporter({ url: 'https://cloud.langfuse.com/api/public/otel' }) });

5. CI/CD Agent Patterns

5.1 CI/CD Components

Purpose: Reusable pipeline configuration units

Structure:

spec: inputs: agent_name: description: "Name of agent to deploy" enable_tracing: type: boolean default: true --- agent-deploy: script: - deploy-agent --name $[[ inputs.agent_name ]] - setup-observability --tracing=$[[ inputs.enable_tracing ]]

Usage:

include: - component: $CI_SERVER_FQDN/agent-platform/components/agent-deploy@1.0.0 inputs: agent_name: "code-review-agent" enable_tracing: true

5.2 CI/CD Steps (Functions)

Purpose: Reusable units within jobs

Example:

agent-execution-job: steps: - function: setup-agent-environment - function: fetch-agent-context inputs: context_sources: ${{steps.setup-agent-environment.outputs.sources}} - function: execute-agent - function: publish-results

5.3 OIDC Authentication

Configuration:

agent-deployment: id_tokens: AGENT_OIDC_TOKEN: aud: https://agent-runtime.example.com script: - authenticate-agent.sh $AGENT_OIDC_TOKEN

Token Claims:

  • Standard: iss, sub, aud, exp, nbf, iat, jti
  • GitLab: project_id, pipeline_id, job_id, ref, ref_protected, user_id

Benefits:

  • Short-lived credentials (5 mins to job timeout)
  • No secret rotation needed
  • Granular access control

5.4 ML/MLOps Integration

Experiment Tracking:

train-agent-model: script: - mlflow run . --experiment-name agent-optimization - mlflow log-metric --run-id $RUN_ID accuracy 0.95

Model Registry:

  • Semantic versioning
  • Up to 5 GB per file
  • Performance metrics, data lineage
  • CI/CD linking for traceability

6. Platform Agents Repository

6.1 Overview

Repository: https://gitlab.com/blueflyio/platform-agents Purpose: OSSA agent registry, manifest conversion, deployment tooling Status: Production-ready with 341 commits, 29 releases

6.2 Canonical Agents (16 Total)

GitLab Domain (4):

  • Merge Request Reviewer
  • Pipeline Remediation
  • Release Coordinator
  • Issue Lifecycle Manager

Other Domains:

  • Orchestration, Validation, Code Quality, MCP, Security, LLM, Documentation, Drupal (3), Infrastructure (2)

6.3 OSSA Manifest Standard (v0.3.2)

Structure:

  • Metadata: name, version, domain, subdomain, type, runtime
  • Access Control: tier-based permissions
  • Capabilities: versioned modules
  • Skills: specialized competencies
  • LLM Configuration: model, temperature, tokens, fallbacks
  • Observability: tracing, metrics, logging

6.4 Bi-Directional GitLab Duo Conversion

Converters:

  • ossa-to-duo.converter.ts - Export to Duo format
  • duo-to-ossa.converter.ts - Import from Duo

14 Total Converters: Claude, Cursor (partial), GitLab Duo, LangChain, CrewAI, AutoGen, Pydantic AI, Dify, n8n

6.5 CI/CD Integration

Agent Suite Framework (ci/agent-suite.yml):

  • Event-driven activation on MR events
  • File pattern-based triggers
  • Bot-based invocation via GitLab API
  • Slash command interface

7. Cursor IDE Integration

7.1 Cloud Agents

Architecture: Autonomous coding assistants in isolated Ubuntu VMs

Capabilities:

  • Asynchronous execution
  • Multi-file code modification
  • Test iteration with auto-fix
  • Branch management
  • CI/CD remediation (GitHub Actions, GitLab planned)

Access Methods:

  1. UI: Cursor: Start Cloud Agent Setup
  2. API: POST /v0/agents
  3. Web: cursor.com/agents

7.2 Cloud Agents API

Launch Agent:

curl -X POST https://api.cursor.com/v0/agents \ -H "Authorization: Bearer $CURSOR_API_KEY" \ -d '{ "task": "Fix all TypeScript errors in src/", "repository": "https://gitlab.com/blueflyio/openstandardagents.git", "ref": "main", "model": "claude-sonnet-4-5", "create_pr": true }'

7.3 MCP Integration for GitLab

Configuration:

// .cursor/mcp.json { "mcpServers": { "gitlab": { "type": "http", "url": "https://gitlab.com/api/v4/mcp" } } }

Available via MCP:

  • List projects, issues, merge requests
  • Create issues/MRs, add comments
  • Trigger pipelines, view status
  • Manage repository files

7.4 Team Collaboration

Workspace Settings:

  • Default model configuration
  • Repository and branch defaults
  • User restrictions
  • Team follow-up permissions

Shared MCP Configs:

  • Define team-wide servers in dashboard
  • Automatic context for all members
  • Custom commands (/review-mr, /fix-ci)

7.5 Cost Management

Pricing (2026):

  • Pro: $20/mo (~225 Sonnet 4 requests)
  • Pro Plus: $60/mo (~675 Sonnet 4 requests)
  • Ultra: $200/mo (~4,500 Sonnet 4 requests)

Combined with GitLab Ultimate (10 devs):

  • GitLab Ultimate + Duo: $1,180/mo
  • Cursor Ultra: $2,500/mo
  • Total: $3,680/mo

ROI: 30-40% productivity gain = 3-4 developer weeks/month


8. Implementation Roadmap

Phase 1: Foundation (Weeks 1-4)

Week 1-2: GitLab Duo Setup

  • Enable Agent Platform in group settings
  • Activate foundational agents (Planner, Security Analyst, Data Analyst)
  • Test foundational flows (Developer, Code Review)
  • Configure AGENTS.md files for project standards

Week 3-4: OSSA Integration

  • Review platform-agents repository
  • Test bi-directional Duo OSSA conversion
  • Define custom OSSA manifests for domain-specific agents
  • Deploy first custom agent via OSSA

Phase 2: Interoperability (Weeks 5-8)

Week 5-6: MCP Setup

  • Deploy GitLab MCP server (expose data)
  • Configure GitLab as MCP client (connect to external tools)
  • Test MCP integration with Cursor, Claude Desktop
  • Build custom MCP server for observability data

Week 7-8: A2A Integration

  • Implement Agent Card at .well-known/agent-card.json
  • Configure agent discovery mechanisms
  • Test agent-to-agent communication
  • Document discovery patterns

Phase 3: Drupal Marketplace (Weeks 9-12)

Week 9-10: Setup

  • Install Drupal 11 with AI modules
  • Configure AI Agents, MCP Server, MCP Client
  • Setup vector database (Milvus or Qdrant)
  • Enable Canvas AI for visual agent building

Week 11-12: Agent Registry

  • Create ai_agent content type with taxonomies
  • Configure JSON:API endpoints
  • Implement ECA workflows for agent chaining
  • Build MCP tools for agent discovery/invocation

Phase 4: Commercial Frameworks (Weeks 13-16)

Week 13-14: Claude Code & OpenAI

  • Deploy Claude Code in GitLab CI/CD
  • Configure OIDC for AWS Bedrock / GCP Vertex AI
  • Integrate OpenAI Agents SDK
  • Setup observability (LangFuse/Helicone)

Week 15-16: Cursor IDE

  • Configure Cursor MCP for GitLab
  • Setup team workspace with shared configs
  • Deploy cloud agents for autonomous coding
  • Integrate API for CI/CD triggers

Phase 5: Production Hardening (Weeks 17-20)

Week 17-18: CI/CD Optimization

  • Create CI/CD Components for reusable agent pipelines
  • Implement CI/CD Steps for fine-grained orchestration
  • Configure SAST, DAST, dependency scanning
  • Setup ML/MLOps experiment tracking

Week 19-20: Observability & Cost

  • Enable distributed tracing across all agents
  • Configure error tracking with alerts
  • Implement DORA metrics dashboards
  • Setup cost attribution and budget alerts

Success Metrics

Technical:

  • Agent response time < 5 seconds
  • 95% uptime for agent services
  • 80%+ agent accuracy on tasks
  • Zero security incidents from agents

Business:

  • 30%+ productivity improvement
  • 50%+ reduction in security vulnerabilities
  • 40%+ faster time to market
  • ROI positive within 6 months

Sources & References

GitLab Documentation

Agent Standards

Drupal Resources

Commercial Frameworks

Internal Resources


Last Updated: 2026-01-07 Research Duration: 10 minutes with 8 parallel agents Total Content: 50,000+ words of comprehensive research Next Actions: Implement Phase 1 of roadmap, expand individual guide pages