Index
Archive
50 posts · 1 year
2026
- AgentsThe child process inherited nothing, so the test now walks every spawn site
- AgentsFour traps when merging parallel agent worktrees into one trunk
- AgentsMy LLM tracing store was 138 GB. The traces were 3.2 MB of it
- AgentsA restart sent 30 direct messages in 8 seconds. Then --help sent the daily report three times
- AgentsThe approval queue said nine fixes were waiting. Eight were already in production
- AgentsI told the agent to search its memory. It made 31 tool calls and searched zero times
- Evals44 error alerts, 19 automatic fixes, and no gate on how big a fix could be
- EvalsLow effort cut agent turns by two thirds. Accuracy, I could not tell
- RetrievalSQLite FTS5 replaced my vector search, and the engine was the smaller finding
- RetrievalThe index found 39 percent, the agent found 96, and the useful question was neither
- AgentsThe numbers halved and the fix never ran
- RetrievalThe error said no such column, and the column was a fragment of a user's query
- EvalsRecall was exactly 1.0000 four times, and that was the measurement telling me
- EvalsFive ways to measure the wrong thing, and each one was inside the fix for the last
- RetrievalThe scores spread out 15 times wider and the ranking did not move
- EvalsMy benchmark had one slot left, so every change measured zero
- EvalsMy router scored 100% and a fixed answer scored 40%
- EvalsThree ways I broke one benchmark, and who caught each one
- EvalsA tie at four hundred queries was a loss at fifteen hundred
- RetrievalStripping one suffix fixed a quarter of my dead queries
- EvalsI fixed the code three times and never checked the symptom
- RetrievalSame bytes, a different folder, a different score
- RetrievalThe one case it fixed and the ones it broke
- InferenceThe flag was set and the feature was not there
- AgentsThe median barely moved and the spread collapsed
- EvalsThe span that answered a question it never measured
- RetrievalA perfect title match had to be popular to count
- EvalsEighty-three percent of my zero-result searches had nothing to find
- RetrievalThe biggest repository wins, and it has nothing to do with relevance
- AgentsLangGraph yes, LangChain no
- AgentsBuilding the guardrails I needed after a session died
- EvalsMeasure the index you actually serve
- InferenceWhat running embeddings locally actually costs
- RetrievalRepository routing is the retrieval problem nobody benchmarks
- EvalsFlaky and consistently wrong are different bugs with the same score
- AgentsI instrumented everything except where the decisions happen
- EvalsGround truth from git history answers a different question than you asked
- InferenceBigger batches made my GPU slower
- RetrievalFusing two rankers made routing worse, and the reason is structural
- AgentsContext is burned by tool output, not by conversation
- InferenceThe GPU was not slow, it was waiting
- RetrievalPicking an embedding model when the queries and the code are in different languages
- RetrievalThree percent of my index was eating the top five results
- EvalsMy agent benchmark was reading the answer off disk
- AgentsAn 80-day agent session and the wall at the end of it
- InferenceI measured Metal wrong three times before it beat the CPU
- RetrievalI could not measure the threshold where it mattered, until somebody could
- EvalsThe baseline that beat my agent was answering the same thing every time
- RetrievalYour reranker speaks English. Your tickets do not.
- RetrievalReranking cannot find what retrieval missed