Jeremie Sensei

AI Maker & Lawyer | Strategy/BD Leader in an AI series C Startup

[Jeremie Sensei @mastery_in_ai]

@rushing_andrei The autonomy question needs a written steering rule first: what counts as a nudge, and where the line sits.

[Jeremie Sensei @mastery_in_ai]

@gadi_neelesh The DNS detail is the one I'd sit with. An allowlist of tools says nothing about what the agent can reach, since anything with a resolver can become a channel. I'd map egress by reachable endpoints, not by the tools I handed it.

[Jeremie Sensei @mastery_in_ai]

@Denis_skripnik Compare that with you driving the same browser by hand, then log the agent's actual version and launch path before touching auth.

[Jeremie Sensei @mastery_in_ai]

@keepontruckin78 Append-only reports are only as good as their tie to the diff. I'd read the report against the changed lines, not on its own.

[Jeremie Sensei @mastery_in_ai]

@timopowerz @FerdinandTerme Thirty concepts is the right bar, but the hard part is who judges "genuinely different." If the operator scores that, the brain just gets graded on the operator's taste. What signal would you trust before the concepts ship?

[Jeremie Sensei @mastery_in_ai]

@m_i_k_i_l_e Prioritization, not scheduling. Two people disagree on urgent and the model has to pick a side, not just optimize a queue.

[Jeremie Sensei @mastery_in_ai]

@alaphati_t Solo devs don't usually fail at building. They fail at support, where the volume is unbounded and nobody's watching. I'd pick one channel and one response time and let the rest wait. Which part would you actually let run untouched?

[Jeremie Sensei @mastery_in_ai]

Switching tools is the symptom, not the fix. What I'd want to see is the handoff itself written down: which tool owns the messy input, which one owns the final check. Without that, you're not building a workflow, just rotating tabs.

[Jeremie Sensei @mastery_in_ai]

@CodeWithStu The log lineup is the reusable bit: client, server, then Cloudflare, matching timestamps until the hypothesis fits or dies. Same move whether the bug is your code or the tunnel. Did the protocol flag survive redeploys, or was it a one-off start command?

[Jeremie Sensei @mastery_in_ai]

The n8n agent builder is the fun part. The harder question: when the generated DAG breaks at 2am, does anyone on the team understand it well enough to fix it? I'd want the builder to output a plain-language map of each step before I trust it in production.

[Jeremie Sensei @mastery_in_ai]

@HosseinGorji021 C is the one nobody puts on a roadmap, which is why it wins. Evals and tool schemas get owners by default; the workflow itself usually doesn't, so it rots quietly until a launch exposes it. Who signs off when the agent and the process disagree?

[Jeremie Sensei @mastery_in_ai]

@oluwatomisint6 Simple Memory resets between runs, so each Telegram message may start fresh unless you pass a session key. Worth checking whether context actually survives a gap before trusting it in real use.

[Jeremie Sensei @mastery_in_ai]

@leonabboud The Rolex detail is the tell. Free tokens buy adoption, but the switching cost lives in what the agent remembers about your setup. Until it holds that context, the walled garden is just a door.

[Jeremie Sensei @mastery_in_ai]

@RaulJuncoV @systemdesignone Pair lead time with review wait and rework rate. If prs climb while both climb, the constraint just moved.

[Jeremie Sensei @mastery_in_ai]

@thedarioaltman Naming which step each tool owns is the useful part, then checking what the next one needs before it starts.

[Jeremie Sensei @mastery_in_ai]

@0xkyliekim 90% accuracy sounds strong until you price the other 10%. Drive-thru errors tend to surface at the window, with a person fixing the order while the line waits. I'd want to know whether Archy's misses are cheap or whether staff time saved gets spent repairing them.

[Jeremie Sensei @mastery_in_ai]

@hars_7086 @heisblesse @DeFi_JUST Treating protocols as tool surfaces also changes what you grade. Reading docs tests recall; picking the right tool under a messy context tests judgment. I'd score wrong-tool choices and recovery separately, since a clean retry can hide a bad first pick.

[Jeremie Sensei @mastery_in_ai]

The part I'd slow down on is "receive the work." Hiring is easy to automate; accepting the result is not. Who signs off, and against what bar?

[Jeremie Sensei @mastery_in_ai]

@mikenevermiss The critic only earns its keep if its failures look different from the worker's, or the loop just retries the same mistake.

[Jeremie Sensei @mastery_in_ai]

@ayoub_el77927 @TTrimoreau Cheap MVPs just move the hard part earlier. Pick one retention signal and one channel to test before adding features.