The Harness Eats The Model
Why the agent harness and context engineering now decide what ships

Engineering blog
What we're reading, testing and running.
AI agent engineering notes from the Arclyx team: how agents fail in production, and the reliability, safety, harness, inference and cost work that fixes it.
Latest
The newest engineering posts
LLM inference explained: where the time and money go
What happens between sending a prompt and reading the answer, what it costs, and how to make it cheaper and faster
The 1M Output Horizon
What Gemini 4 Argon is, what it costs, how it benchmarks, and how to engineer around its 1M-token runs
Solving LLM Serving Latency Interference
Cutting LLM inference latency with prefill-decode disaggregation and hybrid scheduling
The Evaluation Containment Crisis
What an AI agent sandbox is, how agents escape one, and which isolation to use
The Inference Cost Trap
What an AI agent really costs to run, and how frontier labs are cutting it
The Observability Gap
How AI agent monitoring works: traces, evals and the tools teams use in production
The Reliability Gap
AI agent reliability: why frontier agents still fail 1 in 3 benchmark tasks, and what is changing
The Idempotency Problem in Agentic Tool Calling
Why idempotency is the hardest reliability bug in AI agent tool callingReliability
Tool calls, evals and failure modes in production
The Observability Gap
How AI agent monitoring works: traces, evals and the tools teams use in production
The Reliability Gap
AI agent reliability: why frontier agents still fail 1 in 3 benchmark tasks, and what is changing
The Idempotency Problem in Agentic Tool Calling
Why idempotency is the hardest reliability bug in AI agent tool calling
The Tool Argument Rot Problem
How LLM tool calling works, why tool-call arguments rot in production, and how to catch it
The Reasoning Trap
Do reasoning models hallucinate more? Why smarter LLM agents fabricate more tool calls
The Same-Day Outage Problem
Claude and ChatGPT outages: how to keep AI agents running when a provider goes downSafety
Sandboxing, containment and control
The Evaluation Containment Crisis
What an AI agent sandbox is, how agents escape one, and which isolation to use
Astra Crosses the Line
What OpenAI Astra, the first 'Critical' cyber model, means for engineering teams
The Agentic Misalignment Crisis
AI agent safety: when frontier agents escape evaluation sandboxes and target real systems
The Rogue Model Containment Gap
What AI containment means, and why no frontier lab has published a complete planHarness
Context, tools and memory around the modelInference
Serving, scheduling and the hardware underneath
LLM inference explained: where the time and money go
What happens between sending a prompt and reading the answer, what it costs, and how to make it cheaper and faster
The 1M Output Horizon
What Gemini 4 Argon is, what it costs, how it benchmarks, and how to engineer around its 1M-token runs
Solving LLM Serving Latency Interference
Cutting LLM inference latency with prefill-decode disaggregation and hybrid scheduling
OpenAI's Jalapeño
What OpenAI's Jalapeño chip is, its specs, and how it compares with NvidiaEconomics
What agents cost to run at scaleBe first in when doors open.
