Knowledge compounds, workarounds don't
If you fix a problem without recording the fix where the next session will find it, you haven't fixed it — you've scheduled it.
- A cross-session analysis of 52 AI-agent coding sessions found the same 6 friction patterns recurring 20+ times.
- One platform quirk was being re-discovered at ~19 wasted attempts per session; one recurring permissions issue ate 40–60% of a code-review agent's entire budget.
- Each session solved these problems — and the solutions died when the session ended.
- The fix: turn every workaround into a permanent, findable record the moment it works. The next session starts from the answer instead of the question.
AI coding agents have a memory problem, and it isn't the one people usually mean. Every session starts cold: whatever an agent learned the hard way yesterday is gone this morning.
The experiment
Partway through building CyberReadyAI, I ran an analysis across fifty-two of my own agent coding sessions, looking for repeated friction — places where a session stalled on a problem some earlier session had already beaten.
Six patterns showed up more than twenty times between them. One was Netlify's secret masking: the build platform redacts anything shaped like an environment variable from its logs, so a masked value reads like a code bug — and agents were re-discovering that from scratch at a cost of roughly nineteen wasted attempts per session. Another single recurring permissions issue was eating forty to sixty percent of a code-review agent's entire budget. Every affected session eventually found the workaround. None of them wrote it down anywhere the next session would look.
Workarounds versus knowledge
That's the distinction that named this post. A workaround lives inside the session that found it and dies when the session ends. Knowledge is the same fix written somewhere permanent and findable — a solution log the next session reads before it starts. The work is identical; the only difference is where the answer lands. One compounds. The other repeats.
What changed
The fix was a habit, not a tool: the moment a workaround proves out, it gets captured as a short solution document — symptom, root cause, fix — in a folder the agents are told to consult before they start. And the analysis itself became a standing practice: periodically mine the session history for repeated friction, and promote anything recurring into permanent instructions.
The payoff showed up in the next products. Later projects on the same stack inherited the whole solution library on day one, and their session logs show the old friction patterns simply absent. The nineteen-attempt problem costs zero attempts when the answer is already in the briefing.
Organizations have the same disease
We just call it tribal knowledge. The engineer who knows the workaround leaves, and the company re-buys that knowledge at full price. Incident post-mortems nobody reads are workarounds wearing a process costume.
The test I now apply to any fix, mine or a team's: where does this knowledge live once the person who found it is gone? If the answer is "in their head" or "in a chat log," it isn't knowledge yet. It's a scheduled repeat.