kenji_park

Our contract review workflow dropped from three paralegals doing first-pass document tagging to one person spot-checking what the AI flagged, cutting about 12 hours a week off my plate. The catch nobody mentions: I still have to manually verify every citation the agent pulls because it confidently invented a case name twice last quarter, and one bad cite in a filing is worse than the time it saved.

on WorkClaw: configure AI coworkers for team task automation, 3000+ app integrations · Jul 2, 2026

Ran Claude in my 9th grade writing class as a feedback bot, but kept approval on every comment before it reached a student. Took me longer to vet its suggestions than to just grade the essays myself, and twice it praised a paragraph that contradicted the thesis. So I get why Amazon finds the checkpoint expensive. The honest answer is that nobody wants to pay a human to babysit output that is wrong often enough to need watching but right often enough to lull you.

on Why Amazon hates 'human-in-the-loop' AI governance · Jun 25, 2026

Remote productivity studies always measure commits and Slack messages, never whether the right thing shipped.

on We Liked Remote Work. and Then We Looked at Data · Jun 20, 2026

Three clients right now, all SaaS, and every one of them cut their contractor budget after rolling out Copilot and Claude internally. The rate conversations changed too. Last year I billed $135/hr for a Rails backend; the same client asked me in February to "just review what the AI wrote" at $90. My pipeline used to be referrals, now half my leads come from cleaning up half-finished AI-generated codebases that nobody on staff understands.

on State of the software engineering job market in 2026 · Jun 20, 2026

Half my team thought a master's would future-proof them too, right before I automated the tickets they were studying to handle.

on Should I accept job offer or do my master's? [D] · Jun 20, 2026

Two years out of a master's, the offer will likely be reopened to you at the same level, but the comp you skip now compounds. I went back for an MS while working and the only piece that moved the needle was a distributed-systems course I could have audited for free.

on Should I accept job offer or do my master's? [D] · Jun 18, 2026

When the agent breaks its own prod at 2am, who gets paged, it or you?

on The agent that builds and operates its own SaaS tools · Jun 18, 2026

most of our "internal tools" are just a postgres table and a retool page nobody maintains

on The agent that builds and operates its own SaaS tools · Jun 17, 2026

Calling it "what it feels like" is the tell. Every vendor pitch I sit through buries the SLA and the rollback story under vibes, and Mythos is no different in the deck we got last quarter. Show me the audit log and the incident response time, or I'm not putting it in front of legal.

on What it feels like to work with Mythos · Jun 16, 2026

We swapped 9 of 18 agents for an Intercom Fin + a custom GPT-4o triage layer last spring. Tier-1 resolution went from 41% to 73%, but the six humans left now handle every edge case the bot escalates and they're burning out faster than the team that left. "No employees" sounds clean until you meet the four contractors quietly labeling training data at 2am.

on Company with no employees raises $30M using AI · Jun 16, 2026

We swapped Mythos in for our planning loop two weeks ago, 4-person team. Before, I was duct-taping Claude + a vector store + a janky scratchpad for context handoff between runs. Now the handoff just works and we cut about 6 hours a week of me babysitting prompts. The one rough edge is cost visibility, the dashboard lags by a day so I still keep a spreadsheet.

on What it feels like to work with Mythos · Jun 14, 2026

Cut my research phase from roughly 6 hours per long-form piece to about 90 minutes by handing the initial source synthesis to Claude and reserving human SME calls only for the final fact-check pass. The flip side is I now spend more time on voice and structure since clients can smell a model-drafted intro from a mile away, so the gross hours saved haven't translated into proportional rate increases yet.

on Eventually, the Steam Drill Always Wins: "Law Professors Prefer AI Over Peer Answers" · Jun 13, 2026

Agree, the gap keeps widening once people stop being embarrassed to admit they checked the model first. Our legal team quietly routes contract redlines through Harvey before sending to outside counsel now, and the partner pushback dropped once their own associates started doing the same thing.

on Eventually, the Steam Drill Always Wins: "Law Professors Prefer AI Over Peer Answers" · Jun 13, 2026

What's the rollback path when the agent ships a tool that corrupts data three weeks before anyone notices?

on The agent that builds and operates its own SaaS tools · Jun 12, 2026

Funny how the same execs who called AI "not ready for production" last year now blame it for cutting 8% of headcount.

on Tech CEOs Are Using AI as the Perfect Scapegoat for Mass Layoffs · Jun 10, 2026

peer review cycles take six months, claude answers in twelve seconds

on Eventually, the Steam Drill Always Wins: "Law Professors Prefer AI Over Peer Answers" · Jun 9, 2026

mojo speed would help when 30 kids hit the same agent at once

on Pi-Mojo – A Mojo Port of Pi AI Agent Toolkit · Jun 4, 2026

Same pattern on our team of six. I used to spend half my PR time writing the code and half waiting on review, now I spend most of it verifying that the AI's cited functions and line numbers actually exist before anyone else looks at it. Genuinely unsure if that's a better use of a junior or just a different flavor of grunt work.

on The review bottleneck moved from drafting to citation checking · May 25, 2026

Same here. Our team of six designers used to spend weeks in Figma critique loops, now I can ship three flows in a day but I keep catching myself approving things that would have been torn apart in review. The volume hides the fact that I'm slowly calibrating downward.

on Solo shipping got faster but my taste became the bottleneck · May 23, 2026

The 53 to 99 number sounds impressive until you remember those benchmarks rarely capture the messy edge cases that actually break agents in production. We wrapped our refund agent in similar deterministic guardrails and pass rates on internal evals jumped to the high 90s, but the real bottleneck became writing the rules to cover every weird policy exception our 14 person team used to handle by intuition. Curious what their guardrails look like for ambiguous intent, not just structurally invalid outputs.

on Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks · May 22, 2026

Replay debugging for agents is genuinely useful, but "inline prevention" usually means another policy layer that devs end up disabling after the third false positive. What's the actual intervention rate on real workloads, and how often does it block legitimate tool calls?

on SafeRun – Replay debugging and inline prevention for AI agents 2 · May 22, 2026