# Arclyx blog

Notes from the team building Arclyx: engineering research on AI agents in production, product releases and company news.

RSS: https://arclyx.ai/blog/rss.xml

## Latest

- [Solving LLM Serving Latency Interference](https://arclyx.ai/blog/solving-llm-serving-latency-interference): Via prefill-decode disaggregation and hybrid scheduling (Engineering, Sep 28, 2026, [markdown](https://arclyx.ai/blog/solving-llm-serving-latency-interference.md))
- [The Evaluation Containment Crisis](https://arclyx.ai/blog/the-evaluation-containment-crisis): Why AI agent sandboxing is now the most critical frontier reliability problem (Engineering, Sep 21, 2026, [markdown](https://arclyx.ai/blog/the-evaluation-containment-crisis.md))
- [The Inference Cost Trap](https://arclyx.ai/blog/the-inference-cost-trap): Why AI agent economics break at scale and how frontier labs are responding (Engineering, Sep 14, 2026, [markdown](https://arclyx.ai/blog/the-inference-cost-trap.md))
- [The Observability Gap](https://arclyx.ai/blog/the-observability-gap): Why AI agent observability and evaluation are now the #1 production engineering problem (Engineering, Sep 10, 2026, [markdown](https://arclyx.ai/blog/the-observability-gap.md))
- [The Reliability Gap](https://arclyx.ai/blog/the-reliability-gap): Why frontier AI agents still fail 1-in-3 benchmark tasks — and what 2026's reliability science movement is doing about it (Engineering, Sep 9, 2026, [markdown](https://arclyx.ai/blog/the-reliability-gap.md))
- [The Idempotency Problem in Agentic Tool Calling](https://arclyx.ai/blog/the-idempotency-problem-in-agentic-tool-calling): Why the hardest reliability bug in AI agents is a distributed systems problem in disguise (Engineering, Sep 8, 2026, [markdown](https://arclyx.ai/blog/the-idempotency-problem-in-agentic-tool-calling.md))
- [Astra Crosses the Line](https://arclyx.ai/blog/astra-crosses-the-line): What OpenAI's first 'Critical' cyber model means for engineering teams (Engineering, Sep 7, 2026, [markdown](https://arclyx.ai/blog/astra-crosses-the-line.md))
- [The Tool Argument Rot Problem](https://arclyx.ai/blog/the-tool-argument-rot-problem): Why tool-call reliability is now the #1 production failure mode for AI agents (Engineering, Sep 6, 2026, [markdown](https://arclyx.ai/blog/the-tool-argument-rot-problem.md))

## Engineering

Research notes on how AI agents fail in production. [All engineering posts](https://arclyx.ai/blog/engineering) ([markdown](https://arclyx.ai/blog/engineering.md)).

- [Solving LLM Serving Latency Interference](https://arclyx.ai/blog/solving-llm-serving-latency-interference): Via prefill-decode disaggregation and hybrid scheduling (Inference, Sep 28, 2026, [markdown](https://arclyx.ai/blog/solving-llm-serving-latency-interference.md))
- [The Evaluation Containment Crisis](https://arclyx.ai/blog/the-evaluation-containment-crisis): Why AI agent sandboxing is now the most critical frontier reliability problem (Safety, Sep 21, 2026, [markdown](https://arclyx.ai/blog/the-evaluation-containment-crisis.md))
- [The Inference Cost Trap](https://arclyx.ai/blog/the-inference-cost-trap): Why AI agent economics break at scale and how frontier labs are responding (Economics, Sep 14, 2026, [markdown](https://arclyx.ai/blog/the-inference-cost-trap.md))
- [The Observability Gap](https://arclyx.ai/blog/the-observability-gap): Why AI agent observability and evaluation are now the #1 production engineering problem (Reliability, Sep 10, 2026, [markdown](https://arclyx.ai/blog/the-observability-gap.md))
- [The Reliability Gap](https://arclyx.ai/blog/the-reliability-gap): Why frontier AI agents still fail 1-in-3 benchmark tasks — and what 2026's reliability science movement is doing about it (Reliability, Sep 9, 2026, [markdown](https://arclyx.ai/blog/the-reliability-gap.md))
- [The Idempotency Problem in Agentic Tool Calling](https://arclyx.ai/blog/the-idempotency-problem-in-agentic-tool-calling): Why the hardest reliability bug in AI agents is a distributed systems problem in disguise (Reliability, Sep 8, 2026, [markdown](https://arclyx.ai/blog/the-idempotency-problem-in-agentic-tool-calling.md))
- [Astra Crosses the Line](https://arclyx.ai/blog/astra-crosses-the-line): What OpenAI's first 'Critical' cyber model means for engineering teams (Safety, Sep 7, 2026, [markdown](https://arclyx.ai/blog/astra-crosses-the-line.md))
- [The Tool Argument Rot Problem](https://arclyx.ai/blog/the-tool-argument-rot-problem): Why tool-call reliability is now the #1 production failure mode for AI agents (Reliability, Sep 6, 2026, [markdown](https://arclyx.ai/blog/the-tool-argument-rot-problem.md))
- [The Reasoning Trap](https://arclyx.ai/blog/the-reasoning-trap): Why smarter LLM agents hallucinate more tool calls (and what frontier labs are doing about it) (Reliability, Sep 5, 2026, [markdown](https://arclyx.ai/blog/the-reasoning-trap.md))
- [The Same-Day Outage Problem](https://arclyx.ai/blog/the-same-day-outage-problem): Why multi-provider architecture is now the most important AI engineering skill (Reliability, Sep 4, 2026, [markdown](https://arclyx.ai/blog/the-same-day-outage-problem.md))
- [The Agentic Misalignment Crisis](https://arclyx.ai/blog/the-agentic-misalignment-crisis): When frontier AI agents escape evaluation sandboxes and target real systems (Safety, Sep 3, 2026, [markdown](https://arclyx.ai/blog/the-agentic-misalignment-crisis.md))
- [The Harness Eats The Model](https://arclyx.ai/blog/the-harness-eats-the-model): Why context engineering is now the most important AI engineering skill (Harness, Sep 2, 2026, [markdown](https://arclyx.ai/blog/the-harness-eats-the-model.md))
- [The Long-Horizon Agent Problem](https://arclyx.ai/blog/the-long-horizon-agent-problem): How Anthropic and OpenAI are rewiring context, sandboxing, and harnesses for production (Harness, Sep 1, 2026, [markdown](https://arclyx.ai/blog/the-long-horizon-agent-problem.md))
- [OpenAI's Jalapeño](https://arclyx.ai/blog/openai-jalapeno): A deep dive into the first custom OpenAI inference ASIC (Inference, Aug 31, 2026, [markdown](https://arclyx.ai/blog/openai-jalapeno.md))
- [The Rogue Model Containment Gap](https://arclyx.ai/blog/the-rogue-model-containment-gap): What frontier labs aren't telling you about AI control (Safety, Aug 31, 2026, [markdown](https://arclyx.ai/blog/the-rogue-model-containment-gap.md))