// 05 / now
/now
last updated: may 2026
A /now page - what I'm focused on this month. Last updated May 2026.
- Folding agents into GameTools - converting the existing tools into MCP servers with gRPC dispatch, writing packet-analysis skills so the agent can reason about live Blaze traffic to debug new packet shapes, and building the agent loop on top. The infra muscle from the day job lands directly on the side project.
Open To Opportunities
Early-stage startups - collab, freelance, or full-time.
01/ current reading & learning stack
Agent harnesses & agentic loops
anything that pushes on how harnesses get smarter: tool dispatch patterns, context budgets, skill evolution, eval loops, where the boundary between harness and skill should sit. Lately: how teams structure long-running investigations without burning context.
API security
auth flows that fail interestingly, obfuscation, defense in depth practices.
Performance & distributed systems
recent post-mortems and migration write-ups from people who run real load.
Common incidents
engineering blogs where someone explains what broke, why their first guess was wrong, and what they wish they'd had instrumented. The boring ones teach more than the flashy ones.
02/ latest logs & thought vectors
view writing index →Doris json_search was quadratic on arrays. I sent the fix upstream.
A 2 MB JSON document took 18.5 seconds to search in Apache Doris. JSONB arrays have no offset table, and the loop asked for the length and the i-th element on every iteration. The fix is merged into apache/doris.
Finding a psycopg connection leak in executor threads
A pod restarted after the pool hit 50. close_old_connections() ran on the request thread. Each ThreadPoolExecutor worker took a lease and never gave it back.
We removed a column. EU went offline.
DROP COLUMN shipped in the same deploy as the code that stopped reading it. Old pods kept serving traffic against the new schema. EU had zero healthy replicas for about four minutes.