# Research

Engineering and research deep-dives from the team building Viktor.

## AGI harness  
Most agent harness code works around model weaknesses, and workarounds expire with every release. The test behind every Viktor harness decision: will this still be useful for Fable 10?  
July 15, 2026 · [Peter Albert](/content/research/a-harness-for-agi/index.html)

## GPT-5.6 vs Claude: Benchmarked on Real Agent Work  
We ran OpenAI's new Sol, Terra, and Luna against Claude Opus 4.8, Sonnet 5, and Fable 5 on real agent work. The frontier moved — and it has a personality.  
July 9, 2026 · [Toni Albert](/content/research/gpt-5-6-vs-claude-agent-benchmark/index.html)

## How We Built Viktor Around Prompt Caching  
How Viktor's thread engine is designed around prompt caching: SDK instead of tool loading, append-only threads, in-cache compaction, and an 80% cost cut.  
June 8, 2026 · [Toni Albert](/content/research/how-we-built-viktor-around-prompt-caching/index.html)

## What Breaks When Your Agent Has 100,000 Tools  
Most AI agents demo well and fall apart in production. Here's what we learned about tool use, context windows, memory, and UX while building an AI coworker that connects to 3,200+ integrations.  
March 3, 2026 · [Peter Albert](/content/research/what-breaks-when-your-agent-has-100000-tools/index.html)