Paweł · FDE Spot

Building FDE Spot · training for Forward Deployed Engineers
Free Module 1 → · hi@ dies in the meeting? That's the job.

[Paweł · FDE Spot @fde_spot]

Most AI pilots don’t die in training.

They die in the handoff: access, owners, buy-in.

Module 1 of FDE Spot is free → https://t.co/3VYysbVKnS

[Paweł · FDE Spot @fde_spot]

First week on site, before the agent touches prod, I ask the customer one thing: what can it do that we can't undo?

Write that list down. Every item on it gets a human approve step. Everything else can run.

Most teams skip this and find the list during an incident.

[Paweł · FDE Spot @fde_spot]

@Dhanush_Nehru I'd put discovery near the top. Sit with the people doing the job before writing code. That's where you find the bottleneck that actually pays for the pilot.

[Paweł · FDE Spot @fde_spot]

@wholemars Agents opening PRs faster than CI can run them is coming for every team. A per-repo rate limit and a human merge gate are cheap to add before the queue backs up.

[Paweł · FDE Spot @fde_spot]

@Cointelegraph Expected. Agent rollouts at banks end up inside the bank's own network with read-only accounts. The security review is the real timeline, so start it in week one.

[Paweł · FDE Spot @fde_spot]

Before a pilot starts, I write down what would kill it.

Wrong on 1 in 10 real tickets. Ops lead still checking every output after two weeks.

Customers laugh. Then go quiet, because nobody's asked them before.

That list builds more trust than any demo.

[Paweł · FDE Spot @fde_spot]

Before an agent ships, I ask the ops lead:

"When it's wrong, how would you know?"

Usually: "a customer would call."

So we add a queue for calls it's unsure about, plus logs their team reads without me.

Boring work. It's what wins the second contract.

[Paweł · FDE Spot @fde_spot]

@fabiolauria Yeah, agreed. If day one feels clunky they won't open it again, and then the FDE work never even gets a shot. I try to sit with two real users the first day. Catches most of it.

[Paweł · FDE Spot @fde_spot]

@Dachris888 @unclebobmartin Honestly I just ask what's a pain to undo. Writing to their systems, touching money or permissions, anything a customer sees, that gets a hard gate. Reads and drafts I let run and just log everything. If one revert fixes the mistake, I don't make it prove itself every step.

[Paweł · FDE Spot @fde_spot]

Enterprise AI has plenty of tourists. They demo and leave.

After years shipping into messy client tenants, the bar moved. An FDE now builds the workbench: models wired into legacy systems, with security and access sorted first.

[Paweł · FDE Spot @fde_spot]

Week one on an FDE gig, my first ask:

"Can I see the spreadsheet?"

There's always one. Color-coded rows, a tab called FINAL_v3. That's the real process. The SOP doc is the wish list.

Build the agent around the spreadsheet and it survives week two.

[Paweł · FDE Spot @fde_spot]

@sameergkoli Agree. If the model has no card and only ships inside the agent product, leaderboard evals miss what users hit. Grade the harness you actually put in their hands.

[Paweł · FDE Spot @fde_spot]

@the__elliott Would show up for this. Best sessions are usually one messy deploy story, what broke in week two, and the eval that caught it next time.

[Paweł · FDE Spot @fde_spot]

@aamirorbit Real FDE when you own the shipping path into the customer's stack and leave something maintainable. Solutions engineer when it is a deck plus a handoff. Same words, different job.

[Paweł · FDE Spot @fde_spot]

@pauliusztin_ In-harness subagents are great when you want shared permissions and cheap context isolation. Out-of-process starts winning when you need hard blast radius limits on a customer network.

[Paweł · FDE Spot @fde_spot]

@NDTVProfit @tiwary_tanmay The consulting interest is real. The scarce skill is still last-mile work on messy client systems: auth, data ownership, and a stakeholder who can kill the project in one meeting.

[Paweł · FDE Spot @fde_spot]

@ghumare64 Building a harness makes that obvious fast. Context load, tool allowlist, approval gates, and how you resume a failed run are most of the product. The model is rarely the hard part.

[Paweł · FDE Spot @fde_spot]

@Celonis @vasuman @billdetwiler This matches the failure mode we see constantly. A strong model on a broken handoff still burns the week. Fix the process boundary first, then put the agent on it.

[Paweł · FDE Spot @fde_spot]

@ArtificialAnlys The index treating model plus harness as the unit is the right call. In customer deploys we only trust the combo we can actually ship and inspect, not a leaderboard name alone.

[Paweł · FDE Spot @fde_spot]

Anthropic-style FDE loops are not leetcode with extra steps.

Expect stages that probe how you scope a messy customer problem, write a pass bar with them, and defend a miss under pressure.

If your prep is only model trivia, you will fail the room that matters.

[Paweł · FDE Spot @fde_spot]

@AlemTuzlak Custom harness work is where most agent loops fall over. Glad you're calling the agent-loop utility out instead of burying it under chat UX.