Infrastructure Orchestration at Scale - tyga.cloud
Joe Wee · 2026-04-28
The auto-remediation problem
Gartner predicts that by 2029, 70% of enterprises will deploy agentic AI as part of IT infrastructure operations [1]. The promise is clear: AI agents that detect anomalies, diagnose root causes, and fix problems before humans notice. The risk is equally clear: 88% of organisations already using AI-driven remediation report concerns about trust in AI decision-making [2].
The OWASP Top 10 for Agentic Applications [3] identifies two risks that hit infrastructure agents especially hard. ASI08 (Cascading Agent Failures): a single error propagates across connected agents into system-wide impact. ASI10 (Rogue Agents): agents drifting beyond intended scope, making harmful decisions based on corrupted memory or flawed reasoning.
These aren't theoretical. InstaTunnel documented how AI remediation agents treat emergency security patches as "unauthorized drift" and automatically revert them, re-opening the vulnerabilities the patches were meant to fix [4]. Unit 42 built "Zealot," a multi-agent system that autonomously mapped a cloud environment, found misconfigurations, gained credentials, and exfiltrated data within minutes [5]. And Docker Swarm's overlay network has a documented history of FDB table corruption and VTEP failures that can cascade across an entire cluster [6] [7].
The platform
tyga.cloud manages 30+ production Docker Swarm clusters - application clusters, mail clusters, MongoDB replica sets, Redis Sentinel clusters, WebSocket clusters, serverless functions, TURN/STUN relays, and GPU compute. The platform runs 17 SDK modules exposing 112+ agent endpoints for diagnostics, remediation, and deployment.
At the core is the Tyga Orchestrator - an architect/expert orchestration model that runs 60 production workflows (~19K lines of code) with risk-gated step execution, retry/rollback, and real-time SSE streaming.
60 workflows across 3 tiers
Tier 1 (Critical, 8 workflows): Smart stack redeploy (12 steps: compose diff, VTEP fix, crash-loop detection), smart stack deploy (14 steps: DNS + Portainer + ACME), unified cluster deploy (28 steps - works for all cluster types), unified cluster destroy (11 steps), stack force update, mail abuse diagnose, mail abuse remediate, full swarm health diagnose (10 steps with AI verdict).
Tier 2 (Important, 20 workflows): Mail/Mongo/Redis cluster deploy and destroy, mail domain setup (zone + MX + SPF + DKIM + DMARC), Redis cluster with Sentinel + RediSearch + RedisJSON modules, backup create/restore (QA and production), Redis cluster resize, Redis full health diagnose.
Tier 3 (Operational, 25+ workflows): SSL cert management, reverse DNS fix, Traefik config update, Portainer deploy, node provisioning, overlay network fix, FDB surgical cleanup, HA failover test, DNS record fix, OddSockets worker deploy, Tau function deploy, MongoDB collection management, GPU pod management (8 RunPod workflows), stack template discovery and generation.
5 domain expert agents
Expert agents extend a common base class with persistent state, diagnostic compression (using the Tyga Encoder notation format), and self-hosted AI reasoning via your LLM:
| Expert | Domain | Monitors |
|---|---|---|
| Overlay Network | Docker Swarm networking | Stale FDB entries, peer errors, VTEP state across manager nodes |
| FDB Remediation | Forwarding database | FDB accumulation patterns, cleanup thresholds, VTEP fix recommendations |
| App Health | Application monitoring | Task replica health, crash-loop detection, resource usage, service dependencies |
| DNS Propagation | DNS infrastructure | Record propagation tracking, resolution verification, TTL monitoring |
| Base Expert | Foundation | Abstract class: persistent state, Tyga Encoder, LLM AI, safety evaluation |
Risk gates: the critical design decision
Microsoft's Agent Governance Toolkit [8] demonstrated that prompt-based safety has a 26.67% policy violation rate in red-team testing, while deterministic enforcement achieves 0.00%. NIST's AI Agent Standards Initiative [9] calls for least-privilege access and audit trails for agent decisions. Gartner predicts guardian agents will capture 10-15% of the agentic AI market by 2030 [10].
tyga.cloud implements this with a 3-category risk gate system. Every workflow step passes through a risk classification before execution:
| Category | Description | Execution | Examples |
|---|---|---|---|
| Cat 1 | Read-only | Auto-execute | Health checks, metric collection, DNS queries |
| Cat 2 | Safe mutation | Auto-execute (default) | Stack redeploy, service restart, cache clear |
| Cat 3 | Destructive | Requires confirm: true | Cluster destroy, Traefik full redeploy, data migration |
The Tyga Orchestrator enforces these gates before each step. If a Cat 3 step runs without explicit confirmation, it blocks. Failed steps trigger rollback of previously completed steps. This aligns with what Pradhan describes as the core challenge of multi-agent AIOps: "balancing autonomy with oversight to ensure reliability in critical production environments" [11].
Expert agents add a 4-tier safety layer on top: OBSERVE (0), SAFE_AUTO (1), APPROVAL_REQUIRED (2), MANUAL_ONLY (3). Tier 1 actions pass through the A2A TrustGate firewall before execution. The base expert integrates with a2a-safety.js for command evaluation.
Tyga Encoder compression: 94-97%
Infrastructure diagnostics are encoded using Tyga Encoder symbols:
- σ (state) - current metrics:
σ{'{'}eu1|nodes:7|mgrs:3|stacks:58|traefik:HA3{'}'} - δ (anomaly) - deviations:
δ{'{'}peer-errors|sev:HIGH|918 failures in 7 days{'}'} - κ (knowledge) - historical patterns:
κ{'{'}vtepFixRate:0.85|pattern:always_needs_fix{'}'} - ρ (remediation) - actions taken:
ρ{'{'}action:fdb-cleanup|target:mgr1,mgr2,mgr3{'}'} - ψ (history) - timeline:
ψ{'{'}step|ts|duration|result{'}'}
Compression: 7,000+ characters of raw diagnostic JSON to 400-500 characters of Tyga Encoder notation. Your LLM reads the compressed output at 94.7%+ fidelity - enabling self-hosted AI reasoning over infrastructure state on your own infrastructure.
Persistent infrastructure memory
The MongoDB InfraContext model stores persistent state per service instance with 30-day auto-expiry (kept alive on access). This addresses what OWASP classifies as ASI06 (Memory and Context Poisoning) [3] - by scoping memory per service instance, corrupted context in one service cannot propagate to another:
- Service name cache - remembers Portainer service names (no guessing on redeploy)
- VTEP fix patterns - tracks
fixRateto predict if a cluster always needs VTEP repair (>80%) - Deploy history -
avgDeployDurationenables adaptive polling intervals during redeploy verification - Health cache - HTTP response times, TLS certificate status
OCSF audit trail
Every workflow step, every expert finding, every remediation action is logged in OCSF format [13] - the Open Cybersecurity Schema Framework maintained by the Linux Foundation with 900+ contributors and 200+ participating organisations. AWS Security Lake uses OCSF natively. This gives us structured, interoperable audit events that compliance teams can query without parsing custom log formats.
The FDB problem (a real example)
Docker Swarm's overlay network uses VXLAN tunnels with forwarding database (FDB) entries mapping container MAC addresses to tunnel endpoints. A known class of bugs [6] [7] causes stale FDB entries to accumulate - containers that no longer exist still have forwarding rules, causing packets to route to dead endpoints. When enough stale entries accumulate, network performance degrades across the entire cluster.
Without safety gates, a naive auto-remediation agent would flush all FDB entries - including the ones for running containers. That takes down every service on the cluster.
With the risk gate model: FDB diagnostic collection is Cat 1 (auto). The smart FDB cleanup workflow (13 steps) collects live task IPs, cross-references them against FDB entries, and deletes ONLY stale PERMANENT entries - never entries for running containers. Full VTEP reconfiguration is Cat 2 (auto after safety validation). Network failover is Cat 3 (manual only).
References
- Gartner (2025). "Predicts 2026: AI Agents Will Transform IT Infrastructure and Operations." Gartner
- Dark Reading (2026). "Are We Ready for Auto Remediation With Agentic AI?" darkreading.com
- OWASP Top 10 for Agentic Applications (2026). ASI08: Cascading Agent Failures, ASI10: Rogue Agents. genai.owasp.org
- InstaTunnel (2026). "Autonomous IaC Drift: When AI Remediation Reverses Your Security Patches." Medium
- Festinger, Y. and Doytshman, C. (2026). "Can AI Attack the Cloud? Lessons From Building an Autonomous Cloud Offensive Multi-Agent System." Unit 42. Palo Alto Networks
- moby/moby Issue #29024: "Overlay networking VXLAN ARP and intermittent connectivity failures." GitHub
- dhilgarth. "Docker Swarm Network Troubleshooting Guide (FDB Corruption)." GitHub Gist
- Microsoft (2026). "Agent Governance Toolkit." GitHub
- NIST (2026). "AI Agent Standards Initiative (CAISI)." nist.gov
- Gartner (2025). "Guardian Agents will Capture 10-15% of the Agentic AI Market by 2030." Analyst: Avivah Litan. Gartner
- Pradhan, R. (2025). "Multi-Agent Systems in AIOps: Enhancing Detection, Diagnosis, and Remediation." IJCTECE Vol 8, Issue 5. DOI
- Baysal, H. (2026). "Asimov Safety Architecture." IETF Internet-Draft
- OCSF - Open Cybersecurity Schema Framework. Linux Foundation. ocsf.io