Building the Team
With $60M in the bank, the constraint is no longer money: it is people. Gopala has to grow a company of three into a company of about forty in a year, without losing the habits that made Season 1 work: arithmetic before decisions, evidence before claims, and a gate nobody can bypass. This part covers the shape of the organisation, the order of the first 36 hires and why the safety lead is hire number five, how roles are designed, the hiring loop and what research says actually predicts performance, pay and equity, AI agents as teammates, and the hardest part: what Gopala stops doing.
hiring_plan.py and hat_hours.py were run (real output below). Salaries, hire dates, hours and every person are the story's scenario; salary bands are not market data. Research on hiring methods is cited. Gopala and his company are a teaching story. Disclosure: Claude, made by Anthropic, helps write this site.
1. Mission log: months 22โ26
- Draws the target organisation for month 34 before posting a single job; agrees it with the board.
- Research lead and head of infrastructure start (both from his Season 1 network).
- Hires the safety and evaluation lead before any new training work starts; hands over the gate's day-to-day running.
- Headcount 21 at month 26; the first all-hands where Gopala does not know every detail of every project.
2. The organisation at month 34
- Six teams, each owning a layer of Season 1's system map.
- Research owns the model, the data and the self-improvement loop.
- Infrastructure owns the cluster, training and serving platforms (Season 1, Phases 1 and 4).
- Product owns the agents, integrations, the harness and the customer experience.
- Safety and evaluation owns the gate, the evaluation harness, red-teaming and security.
- Go-to-market owns sales, solutions engineering and customer success.
- Operations owns finance, people, legal and compliance.
- Why safety is a separate team.
- In Season 1 the gate was independent because Gopala wrote the rules and a second reviewer checked him. With forty people, the people who propose updates (research) must not be the people who approve them.
- Safety reports to the CEO, not to the research lead, and has a direct line to the independent board director (S2 Part 1).
Who owns what at month 34 (scenario)
From hiring_plan.py: 36 planned hires plus the three people from Season 1.
3. The order of the first 36 hires
- The rule: hire the leads first, then the people they choose.
- Each lead hires and manages their own team; Gopala interviews every hire until month 26 and only the leads after that.
- The first five, and why.
- 1 ยท Research lead (post-training and RL): the loop is the company's core and Gopala can no longer run it alone.
- 2 ยท Head of infrastructure: Maestro-2 needs a bigger machine, and Season 1's on-call rota was two people.
- 3 ยท Head of product engineering: the product has to become a platform (Phase 8).
- 4 ยท Founding account executive: revenue must triple, and founder-led sales does not scale past about ten customers.
- 5 ยท Safety and evaluation lead: before any new research starts producing updates.
- Then the teams fill in.
- Research grows fastest (28% of hires), because the next moat comes from research (Phase 6).
- Go-to-market grows steadily: one seller per three to four engineers is far below a classic software company, because the product is still technical and sold to engineers.
"""hiring_plan.py โ the first 36 hires after the Series A, in order.
Fully loaded cost = salary + equity value is NOT included; this is cash:
salary, benefits, payroll tax, laptop and tools (our assumption: 1.3 x base).
Base salaries are SCENARIO bands for a remote-first company hiring globally,
not market data. Each row: month the person starts, team, role, base $k/yr.
"""
LOAD = 1.3
PLAN = [
(22, "research", "research lead (post-training & RL)", 330), (22, "infra", "head of infrastructure", 300),
(23, "product", "head of product engineering", 280), (23, "gtm", "founding account executive", 190),
(23, "safety", "safety & evaluation lead", 280), (24, "research", "research engineer: environments", 250),
(24, "research", "research engineer: data", 250), (24, "infra", "ML infrastructure engineer", 240),
(24, "product", "product engineer: agents", 220), (24, "ops", "finance & people operations", 170),
(25, "research", "research scientist: RSI levels", 300), (25, "product", "product designer", 190),
(25, "gtm", "solutions engineer", 200), (25, "safety", "red-team engineer", 240),
(26, "infra", "site reliability engineer", 220), (26, "research", "research engineer: evals", 240),
(26, "product", "product engineer: integrations", 210), (26, "gtm", "customer success lead", 170),
(27, "research", "research scientist: pre-training", 310), (27, "infra", "kernel & performance engineer", 280),
(27, "product", "product manager: platform", 220), (28, "gtm", "account executive", 180),
(28, "research", "research engineer: synthetic data", 240), (28, "safety", "security engineer", 240),
(29, "infra", "ML infrastructure engineer", 240), (29, "product", "product engineer: harness", 220),
(30, "research", "research engineer: training", 250), (30, "gtm", "solutions engineer", 200),
(30, "ops", "legal & compliance counsel", 260), (31, "research", "research scientist: agents", 300),
(31, "product", "product engineer: enterprise", 210), (32, "infra", "site reliability engineer", 220),
(32, "gtm", "head of sales", 260), (33, "research", "research engineer: RL", 250),
(33, "safety", "evaluation engineer", 230), (34, "product", "developer relations", 190),
]
def main():
teams = sorted({p[1] for p in PLAN})
print("month headcount people $M/month " + " ".join(f"{t:>8}" for t in teams))
for m in (22, 24, 26, 28, 30, 32, 34):
on = [p for p in PLAN if p[0] <= m]
cost = sum(p[3] for p in on) * LOAD / 12 / 1000 + 0.025 # + Gopala and the two Season 1 hires (scenario)
print(f"{m:5} {len(on) + 3:9} {cost:15.2f} " + " ".join(f"{sum(1 for p in on if p[1] == t):8}" for t in teams))
first5 = [p[2] for p in PLAN[:5]]
print("first five:", "; ".join(first5))
print(f"research share of hires: {sum(1 for p in PLAN if p[1] == 'research') / len(PLAN):.0%}")
if __name__ == "__main__":
main()Headcount and monthly people cost (scenario)
Real output of hiring_plan.py. People cost is cash only; equity is on top.
- What the numbers show.
- People cost grows from about $0.09M to $0.95M a month: by month 34 people cost more each month than the whole company burned at month 18.
- This is the burn line
series_a.pyused in S2 Part 1; if hiring runs 10% over, the board sees it in the same monthly review.
4. Designing roles before posting them
- Every role starts as a one-page brief.
- The outcome the person owns in their first year (for the research lead: "the loop approves at least 12 updates a year without loosening the gate").
- What they will decide alone, what they will recommend, and what they will never decide (S2 Part 4's decision table).
- The three skills that matter most, and how the interview will test each one.
- Titles follow real postings, not invention.
- Roles like ML infrastructure engineer, site reliability engineer, kernel and performance engineer and research engineer match the job families described in the Who Runs the AI Factory series, which grounds each role in real postings and engineering talks.
- The safety and evaluation roles follow the Evaluation Layer and safety evaluations articles.
- Generalists early, specialists later.
- The first infrastructure hire must run training, serving and on-call; the kernel specialist comes at month 27, once there is a training run worth optimising.
5. The hiring loop, and what predicts performance
- What research says works.
- Schmidt and Hunter's 1998 meta-analysis of 85 years of selection research found work samples, general mental ability tests and structured interviews among the strongest predictors of job performance, and unstructured interviews weaker.
- A 2022 re-analysis by Sackett and colleagues corrected for over-adjustment in earlier studies and ranked structured interviews at the top, with most predictors lower than previously thought.
- The practical lesson: ask every candidate the same questions, score them against written criteria, and test the actual work.
- Gopala's loop.
- A 30-minute screen by the hiring lead: the role brief, the candidate's best project.
- A paid work sample, about four hours, on a realistic problem: debug a failing training job, write an evaluation for a skill, or design a gate review.
- Three structured interviews, each scored 1โ4 against the role brief before the interviewers talk to each other.
- Two reference calls asking for specific examples.
- A decision meeting where any interviewer can block, with a written reason.
The hiring loop
Structured, scored independently, and built around the real work.
- What he does not do.
- Puzzle questions, trivia, or "culture fit" without a definition.
- Let the most senior interviewer speak first in the decision meeting.
6. What Gopala stops doing
- The hardest hire is the one that replaces him.
- In Season 1 he wore every hat. Now each hat he keeps is a bottleneck for a team.
- The rule.
- Keep what defines the company: strategy, research direction, the gate's rules, hiring the leads, the board.
- Hand off what someone can do better: coding the infrastructure, on-call, support, most sales calls, finance.
"""hat_hours.py โ where Gopala's 60-hour week goes, before and after hiring.
Hours are the story's SCENARIO estimates from his own calendar review.
The rule he follows: keep what defines the company (strategy, the gate,
research direction, hiring the leads), hand off what someone else can do better.
"""
WEEK = 60
BEFORE = {"coding & infra": 14, "on-call & ops": 8, "customer support": 7, "sales calls": 8, "research": 9,
"the gate (reviews)": 5, "fundraising": 4, "hiring": 3, "admin & finance": 2}
HANDOFF = {"coding & infra": ("head of infrastructure", 12), "on-call & ops": ("SRE rota", 8),
"customer support": ("customer success lead", 6), "sales calls": ("account executive", 5),
"admin & finance": ("finance & people ops", 2)}
NEW = {"hiring": 9, "strategy & board": 6, "1:1s with leads": 5, "fundraising": 2}
def main():
after = dict(BEFORE)
for k, (who, h) in HANDOFF.items():
after[k] -= h
for k, h in NEW.items():
after[k] = after.get(k, 0) + h - (BEFORE.get(k, 0) if k == "fundraising" else 0)
print(f"{'activity':22} {'month 18':>8} {'month 26':>8}")
for k in sorted(set(BEFORE) | set(after), key=lambda k: -after.get(k, 0)):
print(f"{k:22} {BEFORE.get(k, 0):8} {after.get(k, 0):8}")
print(f"{'total':22} {sum(BEFORE.values()):8} {sum(after.values()):8}")
keep = after["research"] + after["the gate (reviews)"] + after["strategy & board"] + after["hiring"]
print(f"share on what only he should do: {(BEFORE['research'] + BEFORE['the gate (reviews)'] + BEFORE['hiring']) / WEEK:.0%} -> {keep / WEEK:.0%}")
if __name__ == "__main__":
main()Gopala's week, before and after (scenario hours)
Real output of hat_hours.py. Share of time on CEO-only work: 28% โ 53%.
- What changes.
- Hiring becomes his biggest single activity: 12 hours a week.
- His week drops from 60 to 45 hours on paper; the rest goes to the unplanned.
- He keeps 2 hours of coding a week, on purpose: reading the code keeps him honest about what the team is building.
7. Try it: hand off a hat
Gopala's week (our model, same numbers as hat_hours.py)
- Things to try.
- Untick the infrastructure handoff: his week goes past 60 hours.
- Tick the research handoff: more time is free, but the share on CEO-only work falls. Handing off research direction would change what the company is.
8. Pay and equity
- Cash.
- Salaries in
hiring_plan.pyare scenario bands for a remote-first company hiring globally; they are not market data. - Fully loaded cash cost is modelled as 1.3 times base (benefits, payroll taxes, equipment), our assumption.
- Salaries in
- Equity.
- The option pool is 12% after the Series A (S2 Part 1). Grants follow a written ladder by level, so two people at the same level get the same grant.
- Standard four-year vesting with a one-year cliff, the common start-up practice.
- The ladder is published inside the company: no private deals.
- Explaining equity honestly.
- Every offer letter shows what the grant is worth at the Series A price, and what it is worth if the company is sold for less than the money invested, when preferences (S2 Part 1) come first.
The pool and the ladder (scenario)
Pool from series_a.py; the split is the story's plan.
9. AI agents as teammates
- Season 1 ran on Gopala plus AI agents. Season 2 keeps the agents, with owners.
- Every agent workflow has a human owner, the same way every service has an on-call owner.
- Where agents do real work (scenario).
- First drafts of evaluation tasks and environment scaffolding, checked by research engineers.
- Triage of alerts and failed jobs, with a human approving any action (Season 1, Part 17).
- Customer onboarding scripts and integration tests.
- Interview logistics and scorecard reminders, never scoring.
- Where they do not.
- Hiring decisions, gate decisions, customer data agreements and anything a regulator might ask about.
- The effect on the plan.
- The plan stops at about forty people, roughly where the story assumes a comparable company without agents would need more; that comparison is an assumption, not a measurement.
10. The first 90 days of the research lead (scenario)
- In week six the research lead proposes merging the gate review into the research team's weekly meeting to save time. Gopala says no, and asks the safety lead to explain why in writing to the whole team.
- What went well.
- The research lead rebuilt the experiment tracker in three weeks and cut the time to run a loop cycle (Season 1, Part 11).
- What went wrong.
- Two research engineers were hired without a work sample because a candidate had "obvious" credentials; one did not work out.
- The lead started approving small updates informally; the gate log showed three updates merged without a second reviewer.
- What Gopala changed.
- The work sample is mandatory for every hire, including referrals.
- The gate tooling now refuses to merge without two recorded approvals (S2 Part 4's
gate_roster.py).
11. Writing down the culture
- Season 1 had habits; Season 2 needs them written.
- Without writing them down, forty people will invent forty versions.
- The six principles (scenario).
- Arithmetic before decisions: every proposal comes with a calculation.
- Evidence before claims: say what was measured, what was simulated and what is a story.
- The gate is not a bottleneck to remove: it is the product.
- Write it down: decisions, reasons and owners go in a document, not a chat.
- Customers' code is theirs: no exceptions, no "just for debugging".
- Disagree in the meeting, commit after it.
- How they are kept alive.
- Each principle has a real story from Season 1 behind it, told in onboarding.
- Performance reviews ask for one example per principle.
12. Remote-first, across time zones
- Why remote.
- The talent for ML infrastructure and RL research is spread across the world; Season 1 already worked this way.
- The rules.
- Two overlapping hours a day for every team.
- Written proposals before meetings; meetings decide, documents explain.
- The on-call rota follows the sun once the SRE team reaches four people.
- The whole company meets in person twice a year.
13. Why the safety lead is hire number five, not twenty-five
- A common mistake is to add safety after the first incident.
- By then, the habits are set, and safety becomes the team that says no to other people's plans.
- What the early safety lead changed.
- Took over day-to-day running of the gate from Gopala, who keeps the rules but no longer reviews every update.
- Started the red-team and evaluation hires before research doubled.
- Wrote the first incident-response drill for the self-improvement loop.
- The trade-off.
- A senior hire with no customers to show for it in the first quarter. The board accepted it because the gate is in the pitch.
14. The first two weeks of every hire
- Day one is planned before the offer is signed.
- A laptop with access set up, a buddy from another team, and a written first-week plan from the hiring lead.
- Week one: the story and the systems.
- Read Season 1's records in order: the idea and kill criteria (Part 1), the plan (Part 3), the run ledger (Part 10), the gate log (Part 11), the scorecard (Part 18).
- Shadow one gate review and one customer call, whatever the role.
- Ship one small, real change by day five: a fixed alert, a test, a document correction.
- Week two: the role.
- A first project with a defined outcome and a check-in at day ten.
- A 30-minute conversation with Gopala about the strategy one-pager (S2 Part 3) and the six principles.
- Why it matters.
- Half the company is less than six months old by month 28. Onboarding is how Season 1's habits survive the arrival of 36 people who were not there.
15. Levels and two career tracks
- A written ladder from the first hire.
- Five levels for individual contributors and three for managers, each described by scope (task, project, team, company) and by examples.
- Titles stay plain: engineer, senior engineer, staff engineer; lead and head for managers.
- Two tracks, equal pay.
- A staff research engineer can be paid as much as a team lead. Nobody has to become a manager to be rewarded.
- Several of Season 2's strongest researchers are expected to stay individual contributors, as is common in research organisations.
- Promotion by evidence.
- A short written case with examples per principle, reviewed by a small committee, twice a year. The manager argues the case but does not decide it.
The two tracks (scenario)
The ladder is the story's; the two-track structure is common practice in engineering organisations.
16. Keeping people
- Why people leave start-ups.
- Unclear direction, a manager they do not trust, no growth, or a better offer. Pay matters, but rarely alone.
- What Gopala measures.
- Regretted attrition: people the company wanted to keep who left, per quarter. Target: zero in the first year, and an honest reason recorded for every case.
- Time from joining to first meaningful ownership.
- A short quarterly survey with the same five questions, reported by team.
- What managers do.
- Weekly one-to-ones that belong to the report, not the manager.
- A written growth plan reviewed every six months.
- What happened (scenario).
- One regretted departure by month 30: a product engineer who joined for the agent work and spent two quarters on integrations. The fix was a rotation rule, not a raise.
17. Eight team-building mistakes
- Hiring for the company of today instead of the one in 12 months.
- Hiring individual contributors before the leads who will manage them.
- Skipping the work sample for an impressive candidate.
- Letting research approve its own updates because it is faster.
- Paying different grants for the same level after a hard negotiation.
- Keeping every hat because nobody else knows how yet.
- Treating agents as headcount instead of tools with owners.
- Writing the culture down only after it has changed.
18. Delegation box
- Hiring the leads and every early hire
- The org design and the culture principles
- The gate's rules (the safety lead runs it day to day)
- Sourcing lists and interview logistics
- Draft role briefs and work samples, checked by the leads
- Running
hiring_plan.pyscenarios for the board
- A recruiting partner for go-to-market roles
- Employment counsel for remote contracts in several countries
- A payroll and benefits provider
19. Budget and risk tracker
| Item | Status after Part 2 | Note |
|---|---|---|
| Headcount | 21 at month 26 โ 39 planned at month 34 | From hiring_plan.py |
| People cost | $0.49M/month at month 26 (scenario) | $0.95M planned at month 34 |
| Gopala's CEO-only time | 28% โ 53% | From hat_hours.py |
| Hiring risk | Watch | One mis-hire in the first ten |
| Gate risk | Lower | Two-approval rule enforced in tooling |
| Culture risk | Watch | Half the company joined in six months |
20. The series
The Maestro, Season 2: 23 parts in five phases (Season 1: 18 parts)
21. Reading the numbers
- Primary / standard
- Schmidt and Hunter (1998) and Sackett et al. (2022) on selection methods; role families from the Who Runs the AI Factory series.
- Secondary โ ๏ธ
- None in this part.
- Our assumptions
- All people, dates, salaries, hours, grants and the 1.3ร loading are the story's scenario.