ENGINEERING BLOG
Technical research and engineering insights from the team building the operating system for responsible AI operations.
121 articles · Published by MARIA OS
From keyword detection to action-level control: a formal shift that recasts AI routing from text classification to governance-aware execution control
Traditional AI routers treat routing as text classification: extract keywords, map to categories, and dispatch handlers. For enterprise-grade agentic systems, this approach is often insufficient. We formalize the Action Router as a function R: (Context × Intent × State) → Action, replacing the naive R: Input → Category mapping. The Action Router integrates with the MARIA OS Gate Engine so responsibility is enforced at routing time, not retrofitted afterward. We formalize the action space, define precondition-effect semantics for routable actions, derive routing cost over feasible actions, and show in simulation that action-level routing reduces misrouting by 67%, cuts responsibility-attribution failures by 94%, and achieves 3.2x lower latency than semantic-similarity routing on enterprise decision workloads.
End-to-end architecture of the three-layer Action Router stack (Intent Parser, Action Resolver, Gate Controller), with recursive optimization and scaling patterns for 100+ agent deployments
The Action Router Intelligence Theory established that routing must control actions, not classify words. This paper presents the full implementation architecture: a three-layer stack of Intent Parser (context-aware goal extraction), Action Resolver (state-dependent action selection with precondition-effect semantics), and Gate Controller (risk-tiered execution envelopes integrated with MARIA OS governance). We detail a recursive optimization loop in which routing policies learn from execution outcomes, formalized as an online convex optimization problem with O(√T) regret. We then present a scaling architecture for 100+ concurrent agents using coordinate-based sharding, hierarchical action caches, and zone-local resolution. Integration with the MARIA OS Decision Pipeline state machine is formalized as a product automaton. Production benchmarks show sub-30ms P99 latency at 10,000 routing decisions per second, with first-attempt accuracy improving from 93.4% to 97.8% after 30 days of recursive learning.
How action routing and gate control compose into a provably safe routing system where each routed action carries complete responsibility provenance
Enterprise AI systems face a core tension: routers must maximize throughput and decision quality, while gate engines must enforce safety constraints and responsibility boundaries. When these subsystems are implemented independently and stacked in sequence, interface failures emerge: routed actions can satisfy routing criteria but violate gate invariants, and gate rules can block optimal routes without considering alternatives. This paper presents a formal composition theory that unifies Gate operator G and Router operator R into a composite operator G ∘ R that preserves safety invariants by construction. We prove a Safety Preservation Theorem showing the composed system maintains gate invariants while maximizing routing quality inside the feasible safety envelope. Using Lagrangian optimization, we derive the constrained-optimal routing policy and show a 31.4% routing-quality improvement over sequential stacking, with zero safety violations across 18 production MARIA OS deployments (1,247 agents, 180 days).
How self-improving routing uses recursive execution feedback to converge toward high-quality policies while preserving Lyapunov stability guarantees
Static action routing — where rules are configured once and applied uniformly — is inadequate for enterprise AI governance. Agent capabilities evolve, workloads shift, and routing quality depends on context that is only observed after execution. This paper introduces a recursive adaptation framework for MARIA OS action routing in which execution outcomes update routing parameters through a formal learning rule. We define θ_{t+1} = θ_t + η∇J(θ_t), where J(θ) is expected routing quality and gradients are estimated from outcome signals. We prove convergence under standard stochastic-approximation assumptions and establish Lyapunov stability guarantees, showing the adaptation process remains bounded while converging toward locally optimal routing policies. Thompson sampling provides principled exploration, and a multi-agent coordination protocol prevents oscillatory conflicts under concurrent adaptation. Across 14 production deployments (983 agents), the framework improves routing quality by 27.8%, converges within 23 adaptation cycles, and records zero stability violations over 1.8 million adapted routing decisions.
How sentence-boundary detection, sequential TTS chaining, and rolling conversation summaries create a natural-feeling voice interface with long-session stability
Voice user interfaces face a core tradeoff: stream tokens immediately for low latency, or wait for larger semantic units to improve naturalness. MARIA OS resolves this with sentence-level streaming: detect sentence boundaries from Gemini token streams in real time, queue each sentence for sequential ElevenLabs TTS playback, and coordinate full-duplex interaction through barge-in control, speech debouncing, and heartbeat-based recovery. This paper presents the cognitive basis for sentence-level granularity, the production `useGeminiLive` architecture, a 29-tool action router across 4 teams with confidence-weighted team inference, and the rolling-summary mechanism for long voice sessions. In 2,400+ production sessions, the system achieved sub-800ms first-sentence latency with zero sentence-ordering violations, including compatibility handling for 9 in-app browser environments.
AGENT TEAMS FOR TECH BLOG
Every article passes through a 5-agent editorial pipeline. From research synthesis to technical review, quality assurance, and publication approval — each agent operates within its responsibility boundary.
Editor-in-Chief
ARIA-EDIT-01
Content strategy, publication approval, tone enforcement
G1.U1.P9.Z1.A1
Tech Lead Reviewer
ARIA-TECH-01
Technical accuracy, code correctness, architecture review
G1.U1.P9.Z1.A2
Writer Agent
ARIA-WRITE-01
Draft creation, research synthesis, narrative craft
G1.U1.P9.Z2.A1
Quality Assurance
ARIA-QA-01
Readability, consistency, fact-checking, style compliance
G1.U1.P9.Z2.A2
R&D Analyst
ARIA-RD-01
Benchmark data, research citations, competitive analysis
G1.U1.P9.Z3.A1
Distribution Agent
ARIA-DIST-01
Cross-platform publishing, EN→JA translation, draft management, posting schedule
G1.U1.P9.Z4.A1
Complete list of all 121 published articles. EN / JA bilingual index.
121 articles
All articles reviewed and approved by the MARIA OS Editorial Pipeline.
© 2026 MARIA OS. All rights reserved.