Ai-Agents

We benchmarked 8 tokenizers from 6 providers and performed exhaustive vocabulary scans. 15 of the most common JSON field names (id, name, type, value, title, time, text, url, path, description) exist as merged vocabulary entries (quote+field in one token) on GPT-4 (#32586 for ‘“name’), LLaMA, and Qwen. Claude and Gemma have zero such entries. These merges are hardcoded, deterministic, and irrecoverable without retraining the tokenizer. On real eval data: JSON boundary merge rate 8.93% vs GCF 1.00% (88.8% fewer). GPT-4 has 114 quote+letter vocabulary entries vs 17 pipe+letter (6.7:1 ratio). JSON overhead is 81% at scale. This structural ambiguity compounds per row and explains model-dependent comprehension failures across 2,400+ evaluations.

23 comprehension runs across 10 models (Claude Opus/Sonnet/Haiku, GPT-5.5/5.4/5.4-mini, Gemini 2.5 Flash/Pro, Gemini 3.1 Pro, Gemini 3.5 Flash). Generation eval across 11 models and 3 providers (Anthropic, OpenAI, Google). GCF wins 22, ties 1, loses 0 on comprehension. GCF achieves 5/5 valid generation on every frontier model with zero prior training. TOON fails 0/5 on generation with Opus, GPT-5.4, GPT-5.4-mini, Gemini 3.1 Pro, and Gemini 3.1 Flash Lite. JSON breaks on input at 500 symbols.

Rigorous benchmark of AI agent code retrieval: 107 tasks, 5 repos, 5 languages, 4 competitors. grep precision: 2%. GitNexus: 7.6%. knowing: 23% (11.5x better, p<0.0001). Plus: 193x faster indexing, 28x less RAM, 48x more token-efficient than Repomix. The first statistically validated comparison of code intelligence tools for AI agents.

AI coding agents have a context problem. The tools solving it fall into four categories: context packers, code graphs, memory systems, and runtime observability. Each solves one piece. None versions the intelligence. None proves anything. None learns without poisoning itself over time. This article explores the landscape and argues that content-addressed code graphs with cryptographic proofs are the missing foundation.

Git proved that content-addressing file contents gives you integrity, history, efficient equality, and distributed collaboration for free. The same architecture applied to code relationships gives you something new: versioned intelligence that you can diff, cache, prove, and trust over time.

We built a reproducible experiment measuring how many tokens AI coding agents consume when navigating code with grep vs LSP. On HashiCorp Consul (319K lines), LSP uses 34x fewer tokens. On a TypeScript rename across 24 files: 1,441x fewer bytes. The experiment covers 4 codebases, 3 languages, 13 tasks covering 7 agent workflows.

MCP servers are the tools AI agents rely on. We tested 55 of them with mcp-assert, found 20 bugs across 9 servers, and submitted fix PRs. Grafana and Ant Group merged ours. Three days after launch, Ant Group’s visualization team asked us to integrate mcp-assert into their CI. The most common failure: servers throw unhandled exceptions instead of returning isError, leaving agents unable to recover.

I needed AI agents to reliably rename symbols, find references, and check diagnostics without silent failures. The existing MCP-LSP tools were stateless, feature-poor, and untested. So I built agent-lsp: a persistent runtime with 50 tools, 20 provider-agnostic skills, speculative execution, and an audit trail for every AI-driven edit.

Agents accumulate autonomous behaviors over time - ‘always do X before Y’, ‘if you see Z then do W’. These instructions eat context budget, drift across invocations, and can’t be observed or tested. How to recognize when an autonomous behavior is a skill waiting to be extracted, and the layered model that makes the boundary clear.

Claude Code agents write code fast. Too fast to catch quality issues in real-time. Here’s how to build validation directly into agent workflows using hooks and team coordination - micro validation after every file write, macro validation before completion, and independent review from validator agents.