← TinyGPT · docs · devlog · roadmap · speedup
source: docs/sessions/2026-06-17-vercel-eve-and-local-agents.md · view on GitHub ↗

Vercel Eve and the local-agent wedge

Date: 2026-06-17 (afternoon). Companion to the morning’s step-back. Captures the strategic decision triggered by Vercel’s Eve launch (blog, public preview June 2026).

What changed today

Vercel shipped Eve — an open-source agent orchestration framework on top of frontier models (Claude primary). Filesystem + TypeScript + markdown config, Workflow SDK for durability, sandboxed exec, OpenTelemetry traces, multi-channel (Slack/Discord/Teams), eval-based CI/CD, multi-agent governance for orgs. Not a model — an infrastructure layer.

Vercel doesn’t ship things at this scale unless they believe a category is forming. The agent-platform thesis is validated at the cloud-infrastructure layer.

The question

User asked: “what about we expand into building ai agents? or too far?”

The answer (and why)

Don’t compete with Eve. From a Mac-first Swift+Python stack, we lose every comparison to the cloud-native incumbent with edge infrastructure and the TypeScript ecosystem. Same lesson as [[feedback_library_api_deferred]] — don’t pitch packaging around nothing.

Do lean into the wedge Eve can’t reach. Eve and every peer (Cursor, Replit Agent, Devin, Cognition) burns frontier-API tokens per invocation. None serves the buyer who needs:

  1. Zero data leaves the machine — regulated industries (legal, healthcare, defense, finance, classified codebases).
  2. Zero marginal cost after training — high-volume internal use on monorepos, batch refactor across services.
  3. Offline operation — air-gapped environments, intermittent connectivity.

tinygpt already owns ~70% of the stack for this wedge:

PieceStatus
QLoRA on 4-14B (tinygpt sft)shipped
Distillation from 30B teachershipped
Trace recorder (B22)PRD only, not yet implemented (corrected 2026-06-17)
Deferred tools in serve (B26)scaffolding shipped, BFCL gate pending
Composite reward framework (B28)scaffolding shipped
Trace-to-training-data (B29)PRD only — blocked on B22
Reasoning-depth classifier (B30)✅ shipped + verified 2026-06-17
Eval CI/CD gate (B32)scaffolding shipped, multi-suite pending
Gallery + project pins (B31)scaffolding shipped

What’s missing is one shipped vertical that proves the loop runs.

The decision

Open [[B35-local-agent-vertical-poc]] as a future kill-or-validate experiment. Code reviewer on a Mac, on a tinygpt-fine-tuned 12B, no cloud. Four-week timebox, explicit kill criterion: if the loop can’t beat zero-shot open baseline by ≥5pp, publish the negative result and keep tinygpt narrow as a model factory.

What this is NOT

A pivot. The decision is to not pivot tinygpt into being an agent company in the Eve sense. The model-factory north star ([[feedback_tinygpt_north_star]]) is unchanged. B35 is one experiment on top of that factory — the factory ships either way.

What’s NOT in B35 scope

Connection to today’s commits

What changes for tinygpt’s day-to-day

Nothing immediate. The factory keeps shipping. B35 sits in §3 Tier B as an explicit option that future sessions can pick up. When the training-data side for code review is ready (PR-review corpus collection), B35 unparks.