dhadrien

I am building towards an orchestration system with a simple north star: going from CEO/CTO to creative director 1: agents execute all 2: agents design solutions 3: agents design strategies then i only learn, collab, get inspired and do creative decisions

hot take: professional subordinated human interactions should disappear. it favors sociopaths coordination/communication should be fair, transparent, via agents and protocols. mentorship, friendship, peership should happen outside of hierarchy, without conflict of interests

good learning loop with agents: 1. all agents with a specific mission, with ACs 2. mission accomplished? launch an assessor subagent with qualitative evals: how well the mission was defined, how efficient the agent, did he reported bugs etc.. 3. nightly processing of findings

if you think claude is bad its either a skill issue or you've been contaminated by psyops def still the best out of there

why building your own harness? worst case: you learn about most important things mid case: you leverage it and import it into the future best harness (still not clear who will win) best case: build something durable ultra customized to you and get an edge over everyone

the most important things is - give access to your agents to things - don't tell him what to do e.g setup MCP with your system commands + add index of tools/skills in their context + some basic global rules I made the mistake to restrain too much my subagents, now they fly

really its day 1, first time im out of specs and out of my claude code, dogfooding the orchestration system

shit it works, agents launch themselves, when complex work notification for plan review, then notification for merge review

problem space exploration of today: how to test agents, and its an interesting one - how to test prompt injection is correct, good context - how to test if he knows what skills to use when - quality assessment of the actual work - integration within deterministic stack

all people releasing skills because it works for them: 1. non ai fluent ppl will misuse and just make their system (openclaw / hermez) worse 2. ai fluent ppl will never use them because vector attack accept that you build things ultra customized to you and yourself only

new mental model for future of software is simple deterministic: spec driven + formally verified agents/llms: eval driven + auto learning i work daily on these problems, hope it'll pay off

or am i just exhausted after 5 hours of deep work?

really thinking about starting to work at 4am in europe, claude way better/faster before noon

claude /compact does work actually quite well, seems like it improved a lot compared to 1 month ago

how to be excited about future agi models vs threatened by them? north star: be the best positioned to make your past work obsolete (it will), when 10x model lands what to work on? on hard problem(s) leveraging edge ai (problems that need learning loops, autonomy, auto-research, agents x ai coordination, agents orchestration etc..) how to build your product? accumulate context and customised workflows - store all raw data (backup all codex/claude/ai transcripts) - document all decisions made: ideally spec driven or some framework to surface explicit decisions made - build knowledge base (rules for agents, learnings, patterns etc..) - be sovereign over your stack how to work with agents? invest in them + you - build ultra customised agentic workflows - custom workflow which encapsulate the essence of how you'd like to work > external more efficient workflow that doesn't really fit how to live with agents? - develop brain athlete mindset (when to work? how much? how? what to do while agents run? workout? etc..) - track deep work - track fun work - find the habits that make your life in general sustainable and better than before ai basic kpis: - how much you learn on your problem + how you work with ai - how many decisions you make - how much fun and peace you get - how long you could sustain your efforts if every 3 months ai does 10 x

had i not my athlete mindset about stacking deep work days, stress, sleep, workout etc.. i think id go crazy with ai im building some crazy powerful shit, ive thrived in deep waters my whole life, got capital to pay for llms, still feeling fomo and anxious

I was .md pilled 1 month ago, but why not just sql + custom frontend?

@nnnnicholas ok thanks for this, doing spec driven and need the opus + 1m context window for most of my days but will def try to use the codex app

@dhadrien_ tbh youll stop using claude code except for run and gunning human in the loop cli stuff -- and even then tbh

regular devs are going to get 2+ codex subs this month

@nnnnicholas ok interesting, so claude might still be the better cli but the codex app so good the switch if worth it for multi sessions management specs in claude, ship sprints in codex app sounds good to you?

@dhadrien_ yes because you're managing so many codex threads the cli paradigm breaks down. even tmux people like steipete have transitioned. because its more of an agent manager and less of a cli at end of day

@nnnnicholas you'd say the desktop app is the game changer? cause I work on ssh and just installed it, it does not seem to support it. is the desktop app way better than cli?

@dhadrien_ download codex app open your home folder or a project folder and create new thread change from Default Permissions to Full Access then talk to gpt 5.4 or codex 5.3. do anything you would have done with claude code