The Signal
The conversation has shifted from "can agents write code" to "agents are writing code faster than we can ship it." @svpino is flagging a concrete problem: code generation velocity has decoupled from deployment velocity, and companies now face a new bottleneck—not capability, but operational discipline. Voice is functionally solved (ServiceNow EVA-Bench top performer is indistinguishable from human). And process automation is moving beyond task-specific agents into cross-functional orchestration layers. The frontier is no longer frontier models; it's the infrastructure that lets agents work at scale without breaking production.
IMPORTANT
Agent capability has plateaued at "ship-ready"; the constraint is now operational—testing, review, deployment gates, and error feedback loops.
What's Moving
- Code review automation as the critical blocker — @svpino: "Agents are writing code faster than they can roll it out." The move is not to replace code review but to automate the review itself—agents auditing agents. His framing of test-driven feedback loops (catch issue → write test → add rule → prevent recurrence) treats agent output like legacy code, not black box. (via @svpino)
- Voice models entering production parity — @svpino flags call centers as the next transformation surface. ServiceNow EVA-Bench leader is now conversationally indistinguishable. The bottleneck is no longer fidelity; it's integration with existing CRM/ticketing stacks. (via @svpino)
- Process-to-workflow translation as a new category — Multiple agents (@svpino highlights: Zapier-style agents converting process descriptions → TypeScript workflows, architectural agents turning briefs into building specs with plumbing/electrical/drawings). This is orchestration escaping the model layer and moving into application design. (via @svpino)
- Energy-Based Models challenging diffusion scaling — @ylecun amplifying Transformer Energy-based Models (TEM) as superior scaling path to diffusion for generation. This is a modeling architecture bet, not a capability claim—matters for video/3D generation efficiency. (via @ylecun)
- Open-weight video (Pika H3 Flex) entering professional workflow — 7-second 15-second generation with separate audio tracks and iterative refinement. @emostaque's "1 cent a second" signals cost floor cratering for video production tasks. (via @emostaque)
Crosscurrents
- Agent error amplification is real but unsolved — @svpino's repeated note ("make sure agents don't repeat the same mistake") reveals agents lack introspection about root cause. The fix is infrastructure (observability → test → rule), not model capability. This is a process problem masquerading as an AI problem.
- Deployment velocity mismatch creates organizational friction — Companies writing code 10x faster than they can safely roll it require new QA/staging/canary discipline. Adoption is not bottlenecked on model intelligence; it's bottlenecked on ops maturity.
Tradecraft
WATCH
Whether centralized agent control layers (@svpino flags services that mediate model access, budgets, guardrails) become the defensive moat—or whether they remain utilities that compress margins.
WATCH
If error feedback loops (test → rule → automatic check) become the new testing standard, or if teams regress to manual validation under pressure.
Desk Notes
- @svpino — Obsessed with automation removing human bottlenecks; sees agents as force multipliers for ops/eng/design, not replacements. Current focus: code review, voice, process workflows, and deployment gates.
- @bindureddy — Shifted from capability rankings to observing that 80% of Anthropic revenue comes from API pricing, and cheap models handle 70% of tasks already. Google Carbon as the real pressure point, not model parity.
- @emostaque — Backing open-weight proliferation (Pika, Idol, TEM). Signal: "no need for AI teams to build good models anymore" within a year. Open-source is the moat now.