OTP-style supervision. Compile-to-determinism. Live time-travel inspection. Behaviour diff. Anthropic prompt caching. MCP-native.
Built for engineers who need agents that don't fall over in prod.
96% of organizations run AI agents. Only 1 in 9 has them in production at scale.
The problem isn't the models. It's that today's agent frameworks were built as scripts: brittle, opaque, expensive, and impossible to debug once they're live. Most failures originate at the boundaries between processes — exactly where current tooling has nothing to say.
Grove takes the lessons from 30 years of Erlang/OTP fault tolerance, adds a workflow compiler, and wraps both in a developer experience that feels like Next.js. It is the runtime layer agents have been missing.
Each one is real, tested, and verified live against Claude.
OTP-style supervisor with one_for_one, one_for_all, rest_for_one. Restart-intensity guard. Let it crash; let the system heal.
Static analysis identifies deterministic paths. Compile prewarms the runtime cache with declared examples. Verified live: 10× projected cost reduction on multi-agent topologies.
Deterministic tools content-hashed by canonical-JSON of their input. Persistent SQLite. LRU eviction beyond 10K entries. Restart your worker, the warm cache stays warm.
Every event recorded by default. Scrub through any past session, click any step, see the full data, fork from any moment. Keyboard scrubbing (←/→/Home/End/F).
Edit agent.ts, save — only the children whose definitions actually changed restart. Siblings keep running.
Declare cases, save profiles, diff across runs. Cases classified same / drift / regressed. Non-zero exit on regression — drop into CI.
Mount tools from any MCP stdio server with one line. Anthropic prompt cache auto-applied to long system prompts; 90% input-token discount on cache reads. Verified live: 4399 tokens cached, 3959 tokens saved on second call.
Targets Bun 1.3+ for native SQLite, native TypeScript, native test, native bundler.
import { agent, supervise, tool } from '@vyntral/grove-core'
import { start } from '@vyntral/grove-runtime'
import { z } from 'zod'
const search = tool({
name: 'search',
description: 'Search the web for a query.',
schema: z.object({ query: z.string() }),
deterministic: true, // ← cache hits cross-process
examples: [{ query: 'sparse attention' }], // ← prewarmed by `grove compile`
run: async ({ query }) => fetchResults(query),
})
const research = agent({
name: 'research',
model: 'anthropic/claude-opus-4-7',
system: 'You research topics rigorously.',
tools: [search],
})
const tree = supervise({
strategy: 'one_for_one',
children: [research],
restart: { intensity: 5, period: 60_000 },
})
const { handle, sessionId } = await start(tree)
const out = await handle.run('survey on prompt cache thresholds')
await handle.stop()Capabilities adjacent frameworks ship today, vs Grove.
| LangChain / CrewAI | DSPy / TextGrad | Replay / AgentOps | Grove | |
|---|---|---|---|---|
| Supervised processes | – | – | – | ✓ |
| Restart strategies (OTP) | – | – | – | ✓ |
| Compile-time optimisation | – | prompt-only | – | workflow-level |
| Persistent runtime cache | – | – | – | ✓ |
| Hot reload (per-child) | – | – | – | ✓ |
| Time-travel inspector | – | – | ✓ | ✓ + fork |
| Behaviour diff (CI-friendly) | – | – | – | ✓ |
| MCP-native (stdio) | partial | – | – | ✓ |
| One install | partial | – | – | ✓ |