Sharing my AI journey
Toronto, Ontario, Canada
Joined September 2026
- Tweets26
- Following45
- Followers12
- Likes10
Chamath says agent teams leak like human teams - with coding agents it's at the subagent handoff. It gets a prompt, but not what went into building it
Like a shift change where the new nurse gets the notes, not the last nurse's experience
What do you use to pass it along?
Harrison says keep the coding agent runs that went well, but I'd evaluate the bad ones too. A failed session shows where it got stuck and what I had to correct
Like a coach who only watches film of the wins
What do you do with your failed runs?
the traces you already log are training data - keep the runs that did the task well, fine-tune a small model on them, eval on your task
a 9B model beating frontier models makes sense when it's trained on exactly your task
smithtune: langchain.com/blog/langsmith…
xitter.cf/LangChain/status/21129…
Coding agents call a task done when the tests go green
That's only one finish line - tests check what someone wrote down, and what I meant lives in my head or an old chat
Like passing inspection with the window on the wrong wall
How do you define done for a coding agent?
Dex says slow down and do less in parallel, but going fast isn't my problem. My problem is I don't have tools that keep me organized once I've got coding agents running across a few sessions
Like a busy kitchen with no order tickets on the rail
What do you use to keep track?
the same thing that has always been true of engineering is still true. Slow down, prioritize well, do less in parallel, elevate getting things #1/2 done over maximizing utilization of your overall attention (even if it means some downtime for you)
Ask a coding agent to fix the same bug twice and you can get two different fixes
That doesn't bother me. What can't vary is why we shipped the one we did
Two plumbers fix one leak two ways - fine, if someone notes which pipe got swapped
What needs to be repeatable for you?
Most people track old agent sessions like parking spots - resume IDs written on whatever's nearby. Helpful until you can't find one. Sessions are valuable; what was tried, what got ruled out and why. We treat it like a receipt for a coffee Where do you track yours?
Most subagents hand back just the answer
Chase asks when a subagent should report back. I care more about what's in it - what it tried, ruled out, and isn't sure of
Like a teammate saying "fixed it" with no word on what they checked
What do your subagents report back?
we are thinking a lot about how to make deepagents more async native. specifically about making async subagents work better/more prominent
one key question is how does a finished subagent "notify" the parent agent? some default behaviors we're considering for when the parent is still running:
1/ interrupt parent run, apply subagent update, continue (basically same as "steer")
2/ wait for parent run to finish, then batch subagent update with any other new messages (basically same as "queue")
another question is how to handle human in the loop? what if main agent is not running but paused and waiting explicit human response to an action?
a/ wait for human to respond, then batch the subagent update with that
b/ resume thread (which would cancel the human interrupt) with the subagent update (thread could then re-ask for the same human approval)
any other options we're missing?
I pick up a coding agent's branch, read the code, and guess the rest
The slow part is rebuilding the decisions - and retrying what already failed
Picture painting a house with no notes on colour choice
What do you wish came with the last agent branch you inherited?
Many people switch machines mid-task, start a new agent session and re-explain. Dex's teleport moves the live session with its context. Picture packing a moving box: what's not inside doesn't arrive. We keep the goal / decisions in the session. What else is in your session?
new @humanlayer_dev release comin at ya
/hl:teleport - move a local session to a ANY remote host you own or launch, without losing context
/hl:orchestrate - use humanlayer to drive humanlayer - split tasks, fork workflows, move artifacts, and more
New minimalist UI w/ rounded corners and way less noise (cmd+K -> minimalist)
Most agent sessions are private chats. When the owner's out, work stalls
Peter's agent lets anyone grab an open session and pings whoever last touched the code
Picture a shared inbox where each ticket names who last touched it
How do you assign work (sessions) on your team?
Used Opus 4.8 for the longest time. Then tried Opus 5 and found it awful; ended turns constantly, lots of mistakes, poor quality designs with fundamental flaws. Switched back to Opus 4.8. Now been using Opus 5.5 for a while and it's quite incredible!
Asking Opus 5.5 for another video, this time an explanation in manga style of how an engineering manager would use capacitor.kurrent.io/ (shared coding session memory across coding agents and your team)
Review flows, because no agent should grade its own homework
As kcap has a remote agent execution capabilities, we used it to add a review flow feature. Ask for an independent review and choose a harness, it spins up a fresh agent, receives a prompt from the implementer, then does the review and reports findings back. Unlike a sub agent, it won’t die at the end. It keeps the context to do consecutive rounds until it reports no findings. I use it all the time for code and specs.
Tony Young retweeted
Every agent memory setup I've tried has added complexity, but still fallen short. Capturing what happened to the code (kinda🫠sorta), but never exactly why.
I don't want a record of what changed. I want the how and why to carry into my next session.
Am I missing something?
Conrad Irwin / Delta: under parallel agents, worktrees break. The thread and its history are what lasts.
I've felt the worktree shuffle: folders multiply, continuity doesn't.
When agents fan out, what do you keep so the next person isn't starting cold?
When a coding agent finishes a session, the next inherits a PR, a summary; not history - a photo not a movie! Lessons are forgotten. Same questions get re-asked. Same mistakes made. What does your teammate inherit when an agent finishes it's work?
I've treated agent sessions as private chat — useful until someone else needs to see or steer what's running
The useful unit isn't my private agent chat. It's a session a teammate can join without a paste-the-transcript retry
How shareable are your agent sessions today?
crazy new drop - you can now use humanlayer directly from Pi agent or @opencode - bring your harness, use it how you like, and make every session collaborative and remote-controllable!
Tony Young retweeted
I don't really know how "local only" people deal with team knowledge. I was taking over AWS Marketplace stuff from a colleague who left the org. Two things would've made my job a lot harder: (1) previous work was done with Claude on a machine that's not mine (2) that session would've been evicted anyway. With kcap.ai I had no issues, just kept asking questions and got everything done within minutes using the prior art that probably took a lot of time and tokens to figure out.
I've handed coding-agent work to a new session mid-workflow, then watched ruled-outs, redirects, don't-try-X fail to follow. Handoff utility is a requirement, not a vibes hope the next agent "just knows." What do you insist travels when the next agent picks up?
Tony Young retweeted
If you've solved something before, especially through working with an agent (no matter which agent), the path to the solution should be kept and easily recalled.
Knowledge should compound. It shouldn't be lost once the task is done.
Anyone else fighting this every session?
wow so i have built like 10 integrations on "GitHub Apps" since it launched in 2018 as an alternative to oauth. I have probably used over 100
none of them was harder to get to the "configure connected repos" than going through chatgpt web to try to connect this to my dot (mind you, this is a page HOSTED ON GITHUB, the provider just needs to link to it)
