How we actually run AI coding agents, so you don't have to learn it the hard way.
We put AI coding agents, including Claude Code and ChatGPT Codex, to real work every day inside our own business and inside client work. Along the way we found real gaps between how these tools are advertised and how they actually behave: instructions that are read but not enforced, context that quietly disappears, rules that only hold if something technical backs them up. This is where we write that down, as we learn it, for anyone else putting AI agents to real work. Each piece below covers one lesson in general terms, then breaks out how it actually plays out in the specific tools we run, tab by tab.
Three lessons worth knowing before you rely on an AI agent for real work.
Each piece stands alone. Together they cover the three places we've seen AI-agent setups quietly fail: not knowing which instructions actually apply, mistaking a written rule for an enforced one, and losing context the moment a session gets long.
Where AI coding agents actually read their instructions from
Global rules, project rules, personal overrides, auto-generated memory, on-demand skills. Every layer gets combined, not replaced, and knowing the order changes how you should write any of them. Covers Claude Code and Codex side by side.
Read the breakdown → EnforcementHooks: how a rule goes from advisory to actually enforced
A written instruction is read and interpreted, not guaranteed. If a rule must hold with zero exceptions, it belongs in an automated check, not just stronger wording in a prompt. Claude Code and Codex handle that check very differently.
Read the breakdown → Session ContinuityHow we keep an AI agent from losing the plot
Long sessions get compacted, and anything that only ever lived in conversation is at risk. A living plan file on disk is what lets an agent pick up cold, days later. Covers Claude Code's auto memory and Codex's session transcripts and memory consolidation.
Read the breakdown →None of this is theoretical. Every practice here came out of a real session where something went wrong first, an instruction that got misread, a fix that silently broke on redeploy, a decision that got lost the moment context ran long, and then got fixed with a technical control instead of a stronger sentence.
We're publishing these as we learn them, not as a finished playbook. If your team is putting Claude Code, or any AI agent, into real production work, this is the series to bookmark.
Rolling this out at your business? Let's scope it together.
We're opening a limited number of scoped engagements: one focused problem in your business, capped at two hours of real work, completely free. If that problem involves putting an AI agent into a real workflow safely, this is exactly the kind of work we do.