Why this is in the vault
Kilo Code's co-founder makes a structural argument — every engineer now manages a fleet of AI agents, which collapses 4-5 traditional roles into one person with a wider mandate — and closes with a claim that analytics/data people are better pre-adapted to that shift than software engineers.
Summary
Emilie Schario is co-founder and head of product/engineering at Kilo Code (open-source, model-agnostic coding agent platform, acquired by Anaconda in July 2026). Tristan Handy interviews her for the dbt Labs podcast. Prior roles: early GitLab data team, ran data at Netlify, data-strategist-in-residence at Amplify Partners, founded ERP startup Turbine (acquired by Settle).
Three load-bearing claims:
"Kilo speed" is a deliberate cultural choice, not an AI-company byproduct. Org of ~20 engineers, exactly one person with "product" in their title. Every other engineer owns their area end-to-end — roadmap, bugs, Discord support — because each is "running a team of AI agents alongside their own work." Her framing: "You don't have a bunch of ICs anymore. You have a bunch of managers, whether they're managing individuals or agents."
Model-agnosticism is a bet on optionality, not lock-in fear. Kilo supports 500+ models, runs early-access programs with most major labs (testing unreleased models for weeks, tuning system prompts, feeding usage data back pre-launch). She draws a distinction from dbt Labs' open-compute argument: for data infra the fear is Oracle-style vendor lock-in; for models "nobody knows what the best model... will be three months out," so staying open preserves task-fit optionality as the frontier moves. Kilo's Agent Manager runs the identical prompt against multiple models in separate git worktrees to compare output.
The junior-engineer problem is unsolved and she's candid about it (avg. engineer tenure 15 years, floor is 10 — a fully remote, seniority-indexed org). Closing argument flips Tristan's question: data people have been "multi-threaded" across marketing/VP/finance asks for years in a way software engineers were protected from — so analytics engineers, already "manager of one," may be better prepared for a world where everyone manages a portfolio of agents than the developers currently building the tooling.
Mapping against Ray Data Co
Ray Data Co's whole design — one founder plus an agent-manager COO layer (Ray) running specialized sub-agents and station pipelines instead of hiring a team — is the exact org shape Schario describes at Kilo, just compressed one step further (0 additional humans instead of "1 PM + 19 engineer-managers"). Her "you don't have ICs anymore, you have managers" reframes the founder's actual day job: dispatch quality (verify-dispatch, station-critic, the fresh-eyes gate pattern already in ~/.claude/skills/) IS the management discipline she's describing, not a workaround for not having a team. The multi-threaded-data-person argument is also a direct validation of the L4→L5 bet (project_l5_north_star_strategic_direction) — the thesis that RDCO's edge is downstream of agent-management capability the founder already built as a data/analytics practitioner, not something to learn from scratch.
⚠️ Sponsorship
This issue is sponsored by dbt Labs, which also publishes the newsletter (house organ, not a paid third party) — includes an embedded dbt Summit 2026 registration pitch (Sept 15-18, Las Vegas, discount code). Bias implication: the interview subject and the interviewer (Tristan Handy, dbt Labs founder) both have commercial interest in framing AI-agent management as validating the dbt/open-data-infra worldview; the "model-agnostic ≠ open-compute" distinction Schario draws is more careful than pure alignment would suggest, but the piece is still promotional context for dbt Summit.
Related
- [[2026-01-20-every-ai-teaching-management]]
- [[2026-02-27-trq212-seeing-like-an-agent]]
- [[project_l5_north_star_strategic_direction]]
- [[feedback_delegation_model_effort_pairing]]