06-reference

technically back to the stone ages platform engineering

2026-11-19·reference·source: Technically·by Megan Reynolds (host; guests Ian Nowland and Camille Fournier)
platform-engineeringcode-reviewcitestingcoding-agentsdelegated-identitycloud-capacity

Why this is in the vault

A podcast-episode digest where two veteran infrastructure leaders (Baseten hardware head, CoreWeave VP Eng) argue that agent-speed coding revives old problems: platform constraints, CI bottlenecks, hollow code review, weak AI-written tests, and agent permissions.

The core argument

Six takeaways from the episode (paraphrased):

  1. A standard platform makes agents better. Fournier argues agents work best against a narrow, known interface; without shared constraints, vibe coding splinters a company into many incompatible solutions. Platform is the common context that keeps agents from drifting.
  2. CI is the new bottleneck. Test flakiness, CI compute cost and who re-enables disabled tests become the constraint; a single centralized pipeline is a choke point.
  3. Most code review is a rubber stamp. Both say review was historically reserved for risky areas, and that now nobody can keep up. Their proposal: risk-rate changes (regulatory, compliance, security, production surface) and require human review only there, with AI scoring doing the triage.
  4. AI-written tests prove little. Agents are satisficers and test what they think they should; heavier end-to-end integration testing becomes the gold standard again, at real compute cost.
  5. The cloud's "server in two minutes" depended on someone planning ahead. Memory shortages and constrained on-demand capacity mean performance engineering comes back.
  6. Delegated identity is the next pain. Agents running with a human's full access, "dangerously skip permissions," and blast-radius questions; sandboxes are described as "better Lambdas" where the agent itself runs.

Mapping against Ray Data Co

Strongest hit: the RDCO fresh-eyes gate (verify-vault-write, sw-critic, behavior-critic, "own work goes to a zero-context critic") is a working answer to takeaways 3 and 4. Fournier's point that an AI tests what it thinks it should test is exactly why the behavior-critic is source-blind and exercises the artifact rather than reading it, and the "risk-rate the change, review where it matters" heuristic matches the tiered REQUIRED-dispatch gating. It also lines up with takeaway 6 against RDCO's own posture: no-secrets-on-disk, auto-mode classifier hard gate, and approval-gated deploys are the delegated-identity controls the guests say the industry lacks. The platform-as-constraint idea supports the skills-and-SOPs-as-shared-context design (brigade stations, CAPABILITIES.md). Mapping is strong for the agent-governance thread, but it reinforces existing practice more than it adds a new decision; no action required.

⚠️ Sponsorship

Railway presents the episode (ad block with $20 credit code, standard Technically sponsor). Per the sender notes Railway is the fixed sponsor of the Vibe Coders series; here it sponsors a non-series interview episode, so it may now be a broader standing sponsor. Bias implication is small: the content is infrastructure-adjacent but the guests (Baseten, CoreWeave) are themselves infrastructure vendors with an interest in platform and capacity narratives.

Related