Tag
#agentic-coding
8 posts
- Agent-infrastructureAgentic-coding-research
Codex Telemetry Shows the Agent Is a Runtime, Not a Chat Window
A Codex telemetry overlay with 731,341 logs and 1,180,462 spans shows where agent work actually lives: streams, turns, tool dispatch, and long-tail latency.
Michael Isaac - Agent-infrastructureAgentic-coding-research
Stuckness Is Where Agentic Coding Gets Expensive
In my Claude Code corpus, sessions with 10 or more error signals represented 49.7% of token-bearing sessions but 97.1% of token volume. The lesson is not avoid errors. It is interrupt stuckness.
Michael Isaac - Agent-infrastructureAgentic-coding-research
The Agent Loop Map: Claude Code Is Mostly Search, Edit, Shell, Repeat
In 247,592 Claude Code tool events, 81.7% of tool-family transitions stayed inside the search/edit/shell core. The leverage is not just prompt engineering. It is loop-shaping.
Michael Isaac - Agent-infrastructureAgentic-coding-research
Autonomy Has a Half-Life: What 247,592 Tool Calls Say About Claude Code Checkpoints
Real agentic coding is not infinite autonomy. In my Claude Code corpus, nonzero agent runs between human messages had a median of 5 tools and a p90 of 25. That points to a concrete checkpoint protocol.
Michael Isaac - Agent-infrastructureAgentic-coding-research
Claude Code Verification Debt: The Agent Said Done, But Where Are the Receipts?
In 450,878 Claude Code assistant turns, 90.2% of completion-claim turns landed in a same-turn unverified candidate bucket. That does not mean the work was wrong. It means operators need receipts.
Michael Isaac - Agent-infrastructureAgentic-coding-research
fpk: F-Bombs Per Thousand. The Dev-Experience Metric You Didn't Know You Needed
I scanned 5 months of my own Claude Code conversation logs for f-bombs and correlated the rate with model and CLI version. The result was a surprisingly clean DX gradient, and a metric I'm only half-joking about.
Michael Isaac - Agent-infrastructure
Codex vs Claude Code: Which Wins for Production Agent Work (2026)
I shipped production code with both Codex and Claude Code for six months. Honest comparison of cost, autonomy, plugin ecosystems, and the failure modes nobody mentions.
Michael Isaac - Agent-infrastructureAgentic-coding-research
What I Learned From 245,306 Claude Code Tool Calls
I analyzed 245,306 Claude Code tool calls across 113 days. The data says it's a Unix operator with an LLM loop, not a chat product with tools attached.
Michael Isaac