greg isenberg spent 48 hours in sf with ai founders and execs and wrote down 21 things he noticed. i build rema, which onboards your coding agent to your team. 8 of his 21 are about the problem i work on every day: what an agent knows about the place it works.
the model is the part that changes
his note 4 is the frame for the rest. "The model is the CPU, the harness is the operating system", and skills and plugins are the apps. note 6 says the best teams treat the model as swappable, and that in a vertical harness "The model is maybe 20% of it."
i agree, and it's why rema is a plugin. it's one hook that runs inside the agent you already use. in september the same hook file ran unchanged in claude code and in codex.
note 2 is the question someone asked at dinner: when a new model drops next month, "does your company get better or worse?" for team context the answer is better. the passages rema hands the agent don't change. the model reading them gets smarter.
a smarter model also doesn't fix what nobody told it. in september i asked opus 5.5 a real question about a team's uv config. without the team's threads it got the meaning of the setting wrong 4 times out of 4. with them it got it right 4 of 4. the answer lived in a review thread and an issue reply, and no model has read those for your team.
piping it all in isn't onboarding
note 10 is the one i felt most. every company he talked to wants its internal data in one place, exposed to agents over mcp. then they look at it and find "five versions of the same customer, old Slack channels" and pdfs nobody can search.
a new engineer doesn't learn a team by reading every slack channel either. someone tells them what was decided and points to where it's written down. that's the job rema does for the agent. it puts what your team already worked out in front of it, and what you can install today puts the source and file date on every passage, so a stale one can be caught.
stale is the real risk here, and i've hit it. the first time rema ran inside a real agent, it handed over a uv error report without the maintainer's reply one comment down, and the agent told the user the error still happens. a hit now brings the next comment in its thread with it.
failures and memory
note 12 says labs can get training data, but "real examples of agents failing inside companies" are hard to get. inside a team those failures are mostly dead ends: what an agent tried, what didn't work and why. case 03 is where i started keeping them as data next to the fix.
note 13 is a consumer exec saying "the next fight is over memory". note 5 is an agent that can "watch you work for a week and then do your job". both are about one person. a team's knowledge isn't on one person's screen. the decision was in a review thread someone else wrote, and the fix was in another engineer's session. faster developers don't make a faster team when that context stays siloed. rema keeps a team's notes on each engineer's own machine today, and the shared half is in review.
note 19 is the hire everyone wants: "someone with taste who can write evals". every number on this blog comes from an eval i wrote before i trusted the result. case 04 is the longest one.
| note | what he noticed | where rema is |
|---|---|---|
| 2 | a new model makes some companies worse | the same passages go to a better reader |
| 4 | model, harness, then skills and plugins | a plugin: one hook in the agent you use |
| 5 | an agent that watches you work for a week | captures your agent sessions |
| 6 | the model is swappable | the same hook ran in claude code and codex |
| 10 | internal data is a mess | source and file date on every passage |
| 12 | failure logs are valuable | dead ends kept next to the fix (case 03) |
| 13 | the next fight is memory | per machine today, shared half in review |
| 19 | taste plus evals is the hire | every number here comes from an eval |
if you hit this
ask greg's dinner question about your own setup. if a new model ships next month, what will your agent still not know about your team? write that list down first, with where each answer lives.
what this changed in rema
the model is the part that keeps changing. rema keeps the part that doesn't, what your team already worked out, with the source on every passage, for whichever agent you run.
