Skip to main content

Autonomous Enterprise Agent — Search Strategy & Landscape Analysis

Date: 2026-02-19 Author: Claude (Opus 4.6) Project: PILOT Goal: Build an autonomous bot that extends coditect-core beyond business/development into enterprise system integration (Gmail, calendars, desktop publishing, document management, etc.)


1. Strategic Intent

CODITECT currently operates as a development and business platform. The next evolution is an autonomous enterprise agent that can:

  • Discover enterprise systems (Google Workspace, Microsoft 365, desktop apps)
  • Operate them autonomously (read/send email, schedule meetings, create documents)
  • Integrate data flows back into coditect-core (session logs, analytics, project management)
  • Secure all operations with audit trails, sandboxing, and permission models

License Compatibility Requirements

LicenseCompatibleNotes
MITYesMaximum freedom, preferred
Apache 2.0YesPatent grant included, excellent for enterprise
BSD (2/3-clause)YesMinimal restrictions
ISCYesEquivalent to MIT
MPL-2.0YesFile-level copyleft only, acceptable
LGPL-2.1/3.0ConditionalOK if used as library, not modified
AGPL-3.0NoNetwork copyleft — process boundary separation required
GPL-2.0/3.0NoFull copyleft — cannot integrate into proprietary
Fair-code (n8n)ConditionalSelf-host OK, redistribution restricted
BSL/SSPLNoNot truly open source

2. Search System Prompts

Wave 1: Core Agent Frameworks

PROMPT 1 - General Agent Frameworks
"open source autonomous AI agent framework enterprise integration 2025 2026 MIT Apache license github"

PROMPT 2 - Multi-Agent Orchestration
"open source multi-agent AI framework tool use MCP model context protocol 2026 comparison"

PROMPT 3 - Enterprise Workflow Agents
"open source AI agent Gmail calendar Slack enterprise system integration agentic automation framework"

PROMPT 4 - Framework Comparison
"best open source AI agent frameworks 2026 comparison autonomous tools MCP LangGraph CrewAI AutoGen"

Wave 2: Desktop & Browser Automation

PROMPT 5 - Desktop Agents
"open source computer use agent desktop automation browser agent 2025 2026 github MIT Apache"

PROMPT 6 - GUI Automation
"open source AI GUI agent screenshot automation desktop Linux macOS enterprise 2026"

PROMPT 7 - Browser Agents
"browser-use Playwright autonomous web agent open source MIT license Python automation"

PROMPT 8 - Containerized Desktop
"containerized AI desktop agent Docker Linux virtual desktop automation open source self-hosted"

Wave 3: Specific Projects Deep-Dive

PROMPT 9 - OpenClaw
"OpenClaw open source AI assistant MCP integration license 2026 enterprise automation architecture"

PROMPT 10 - Accomplish AI
"accomplish-ai accomplish github open source desktop agent framework Electron MIT license"

PROMPT 11 - Bytebot
"Bytebot open source desktop agent Apache license containerized Linux automation Docker 2026"

PROMPT 12 - Agent S2
"Agent S2 simular-ai GUI agent open source license enterprise desktop automation framework"

PROMPT 13 - Browser Use
"browser-use browser agent open source MIT license Python autonomous web 60k stars"

PROMPT 14 - Skyvern
"Skyvern AI browser automation open source AGPL license enterprise workflow 2026"

PROMPT 15 - IBM CUGA
"IBM CUGA configurable generalist agent open source Apache license enterprise workflow MCP"

Wave 4: Security & Architecture Patterns

PROMPT 16 - Agent Security
"AI agent security best practices enterprise sandboxing permission model 2025 2026 autonomous OWASP"

PROMPT 17 - Agent Isolation
"AI agent sandboxing microVM gVisor container isolation security production 2026"

PROMPT 18 - Permission Models
"autonomous AI agent permission model human-in-the-loop approval gate enterprise security"

PROMPT 19 - MCP Security
"MCP model context protocol security authentication authorization enterprise tool integration"

PROMPT 20 - Audit & Compliance
"AI agent audit trail compliance logging enterprise autonomous action tracking immutable"

Wave 5: Ecosystem & Integration Patterns

PROMPT 21 - Google Workspace
"AI agent Google Workspace Gmail Calendar Drive integration open source API automation"

PROMPT 22 - Microsoft 365
"AI agent Microsoft 365 Outlook Teams SharePoint integration open source Graph API"

PROMPT 23 - Desktop Publishing
"AI agent document creation PDF Word Excel automated desktop publishing open source"

PROMPT 24 - MCP Ecosystem
"MCP server ecosystem open source tools integrations catalog 2026 model context protocol"

PROMPT 25 - Workflow Orchestration
"n8n Temporal Prefect open source workflow orchestration AI agent integration self-hosted enterprise"

Wave 6: Architecture & Patterns

PROMPT 26 - Agent Architecture Patterns
"autonomous AI agent architecture patterns tool use planning memory state management 2026"

PROMPT 27 - Agent-to-Agent Communication
"multi-agent communication protocol A2A Google agent-to-agent open source 2026"

PROMPT 28 - Agent Observability
"AI agent observability monitoring tracing LangFuse LangSmith open source 2026"

PROMPT 29 - Local-First Agents
"local first AI agent privacy enterprise on-premise self-hosted open source desktop"

PROMPT 30 - Agent Evaluation
"AI agent evaluation benchmark testing framework open source autonomous reliability"

3. Initial Landscape Analysis (Wave 1-3 Results)

Tier 1: Primary Candidates (License-Compatible, High Maturity)

ProjectLicenseStarsCategoryMCPEnterprise Ready
OpenClawMIT140K+Personal AI AssistantYes (server)Medium — moving to foundation
AccomplishMITActiveDesktop AI CoworkerVia OpenCodeMedium — Electron + React
Browser UseMIT60K+Web AutomationPlannedHigh — 89% WebVoyager
BytebotApache 2.0ActiveContainerized DesktopVia toolsHigh — Docker isolated
CrewAIMITHighMulti-Agent OrchestrationYes (native)High — 500+ integrations
LangGraphMITHighAgent State MachineYes (adapter)High — LangChain ecosystem
AutoGenMIT40K+Multi-Agent ChatYes (extension)High — Microsoft backed
IBM CUGAApache 2.0NewEnterprise WorkflowYes (native)High — IBM enterprise focus
Agent S2Apache 2.0ActiveGUI Desktop AgentNoMedium — research-grade
Semantic KernelMITActiveEnterprise Agent SDKVia pluginsVery High — auth/audit built-in

Tier 2: Conditional Candidates (License Concerns or Partial Fit)

ProjectLicenseIssueMitigation
SkyvernAGPL-3.0Network copyleftProcess boundary separation only
n8nFair-codeRedistribution restrictedSelf-host OK, cannot bundle
Crawl4AIApache 2.0Scraping only, not full agentUseful as component
ProjectLicenseReason
Sidecar (CodeStory)AGPL-3.0Already analyzed — process boundary only (ADR-165)
Any GPL-3.0 projectGPL-3.0Cannot integrate into proprietary platform

4. Evaluation Criteria Matrix

Rate each candidate 1-5 on these dimensions:

DimensionWeightDescription
License Compatibility25%MIT/Apache > MPL > LGPL > AGPL/GPL
Enterprise System Coverage20%Gmail, Calendar, Drive, Office, desktop apps
Security Architecture20%Sandboxing, permissions, audit trails, HITL
MCP/Tool Integration15%Native MCP support, extensible tool system
coditect-core Alignment10%Architecture fit with hooks, agents, skills, commands
Community & Maintenance10%Active development, stars, contributors, funding

Preliminary Scoring

ProjectLicense (25)Enterprise (20)Security (20)MCP (15)Alignment (10)Community (10)Total
OpenClaw2516121281083
Bytebot25161896680
CrewAI251614156884
Browser Use251210961072
Accomplish25141498676
Semantic Kernel251620124885
LangGraph251412156880
IBM CUGA251416156480
AutoGen251214126877
Agent S225101464665

5. Security Architecture Requirements (OWASP Agent Security Top 10)

Based on current best practices (2026):

RiskMitigation Required
Excessive AgencyLeast-privilege permissions per tool; explicit action scoping
Prompt InjectionInput validation, output filtering, system prompt isolation
Insecure Tool UseTool allowlisting, parameter validation, rate limiting
Data ExfiltrationNetwork sandboxing, outbound connection allowlists
Privilege EscalationRole-based access, temporary scoped tokens
Insufficient LoggingImmutable audit trails for all agent actions
Uncontrolled AutonomyHuman-in-the-loop gates for high-risk actions
Supply ChainSigned artifacts, dependency scanning, SBOM
Model ManipulationInput/output guardrails, content safety filters
Denial of ServiceRate limiting, resource quotas, circuit breakers
  1. Sandboxing: Docker containers (Bytebot pattern) or microVMs for untrusted code
  2. Permission Model: coditect-core hooks as approval gates (PreToolUse pattern)
  3. Audit Trail: Extend session logging to capture all agent actions
  4. Network Isolation: Allowlisted outbound connections per tool
  5. HITL Gates: CoditectModal approval for destructive/financial/external actions
  6. Token Scoping: OAuth2 scoped tokens per integration, auto-expiring

Rather than adopting a single framework, build a layered architecture:

+------------------------------------------+
| CODITECT Enterprise Agent |
| (coditect-core agents/commands/hooks) |
+------------------------------------------+
| Orchestration Layer |
| CrewAI or LangGraph (multi-agent) |
+------------------------------------------+
| Execution Engines (pluggable) |
| +----------+ +----------+ +-----------+ |
| | Browser | | Desktop | | API/MCP | |
| | Use | | Bytebot | | Direct | |
| +----------+ +----------+ +-----------+ |
+------------------------------------------+
| Security & Observability |
| Sandboxing | Audit | HITL | Monitoring |
+------------------------------------------+
| Enterprise Integrations (MCP) |
| Gmail | Calendar | Drive | Office | ... |
+------------------------------------------+

Component Selection

LayerPrimary ChoiceRationale
OrchestrationCrewAI (MIT)Highest score, native MCP, multi-agent
Browser ExecutionBrowser Use (MIT)60K stars, 89% accuracy, MIT
Desktop ExecutionBytebot (Apache 2.0)Containerized, isolated, enterprise
API/MCP ToolsSemantic Kernel (MIT)Enterprise auth/audit built-in
ObservabilityLangFuse (MIT)Already in coditect ecosystem
Workflow Fallbackn8n (self-hosted)400+ integrations, visual builder

7. coditect-core Integration Points

coditect-core ComponentIntegration
Hooks (118)PreToolUse/PostToolUse for agent action approval gates
Agents (776)New agent type: enterprise-system-agent
Commands (377)/enterprise-agent start, /enterprise-agent status
Skills (445)Enterprise automation skills (Gmail, Calendar, etc.)
Session LoggingAll agent actions logged to session log
MCP ServersExtend existing semantic-search, call-graph, impact-analysis
Ralph WiggumRun enterprise agents as autonomous loops with checkpoints
Context DBStore enterprise system state in org.db decisions table

8. Next Steps

  1. Deep-dive top 3 candidates (CrewAI, Bytebot, Semantic Kernel) — clone repos, analyze architecture
  2. Add to coditect-bot — already started (Accomplish added as b500d9a)
  3. Prototype MCP bridge — connect coditect-core MCP servers to enterprise agent
  4. Security architecture ADR — draft ADR for enterprise agent sandboxing model
  5. Run Wave 4-6 searches — security patterns, integration specifics, architecture deep-dives
  6. Create TRACK tasks — formalize as H-track or new dedicated track

Sources

Agent Frameworks

Desktop & Browser Agents

Enterprise Integration

Security