1. What Each Product Actually Is
TypeSafe AI — Jev
Founded by Diogo Almeida (formerly OpenAI, co-creator of InstructGPT and the RLHF methodology behind ChatGPT), TypeSafe AI launched Jev in September 2026 after two years in stealth. Jev is a System One Model — a purpose-trained model for fast, structured, probabilistic decisions that software can consume directly.
Jev answers three question types and nothing else:
| Primitive | Returns | Confidence? | Typical Use |
|---|---|---|---|
| Choice | One option from a defined set + probability per option | Yes (0–1) | Classification, routing, categorisation |
| Score | Numeric value across ordered levels + probability per level | Yes (0–1) | Sentiment, severity, quality, risk |
| Noul | Probability that a yes/no statement is true (0.0–1.0) | No | Binary checks, guardrails, verification |
No strings. No free text. No code generation. No tool calls. No memory. No agent loop. The schema defines the entire output space. The model uses a parallel sampler — evaluating probability distributions in a single forward pass rather than generating tokens sequentially — achieving 70–500ms end-to-end latency.
Myelix Agents
Myelix Agents is a fault-tolerant multi-agent orchestration runtime for autonomous software development. Built in Elixir on the BEAM VM, it orchestrates swarms of AI agents with:
- Isolated Git Worktree workspaces per agent (instant, no container overhead)
- Persistent AST code intelligence graph via Synapse (50+ languages, local, cross-session)
- OTP supervision trees — microsecond crash recovery, true BEAM concurrency
- 5-stage satisfaction evaluator with structured gate violation feedback
- 3-tier prefix-cached prompt architecture (85–91% LLM cost reduction)
- Adaptive neural network critic (NxCritic LSTM) learning from execution trajectories
- CommandPolicy Elixir AST filtering before any shell execution
- Confidence-aware autonomy tiers and per-tool human approval gating
- Steer directives — mid-run developer corrections without interrupting execution
- VWO (Virtual Worktree Overlay) — Synapse results for child-agent worktrees shadow parent chunks correctly via repo_id scoping
- Degraded Synapse inventory — transparent fallback to local grep/file_glob when Synapse unavailable, no action required from agents
- Planning phase gating — Synapse serves read-only tools during planning; write tools only unlocked in execution phase
- Playwright browser automation as a native tool
- Stateful PTY terminal emulation with interactive programme support
2. The Stack Positions
MYELIX AGENTS
Orchestration runtime layer — Multi-step, stateful, fault-tolerant agent loops. PM coordinator decomposes tasks into a kanban; sub-agents execute concurrently in isolated worktrees; gates verify before merge; NxCritic learns from trajectories.
INTEGRATION SEAM
Jev called as a Cortex tool — Intent routing at <100ms, diff risk scoring as Noul questions, confidence-gated autonomy tier escalation. The most natural and powerful integration point.
TYPESAFE JEV
Decision primitive layer — Typed, calibrated, probabilistic function calls. Schema-enforced output space. 70–500ms latency. $0.042/MTok. Stateless, no memory, no agent loop — by design.
LLM LAYER
Frontier models (Anthropic, OpenAI, etc.) — BYOK, multi-model routing. Both platforms route to frontier LLMs for generative tasks. TypeSafe replaces the LLM for classification; Myelix Agents orchestrates the LLM for generation.
3. Feature Comparison Matrix
3.1 Core Capabilities
| Capability | Myelix Agents | TypeSafe / Jev |
|---|---|---|
| Multi-step agent loop | ✅ Full planner loop, PM coordinator, subagents, round budget | ❌ Deliberately absent — stateless function calls |
| Code intelligence (Synapse graph) | ✅ Persistent structural AST graph — 50+ languages, cross-session, local | ❌ None — caller assembles context manually |
| Structured output | 🟡 Enforced via action schema IR + JSON parsing + gate system | ✅ Native schema-enforced — mathematically guaranteed by output space closure |
| Calibrated confidence scores | 🟡 Gate violations as pass/fail; autonomy tiers as risk bands; NxCritic trajectory scores | ✅ Native — probability distribution per option, mathematically derived confidence scalar |
| Hallucination prevention | ✅ Pre-write Synapse simulation + 5-gate evaluator + CommandPolicy AST filtering | ✅ Output-space constraint — schema-enforced, zero hallucination on type (wrong-but-valid still possible) |
| Context management | ✅ 3-tier tiered prompt cache, relevance-aware pruning, Synapse structural grounding, VWO overlay | ❌ Manual — developer responsibility to assemble correct context |
| Persistent memory | ✅ Session learnings, evidence ledger, clarification log, NxCritic trajectory store | ❌ Stateless — no memory between calls |
| Workspace isolation | ✅ Git Worktrees per agent — instant, no container overhead | ➖ Not applicable |
| Fault-tolerant processes | ✅ BEAM OTP supervision trees — microsecond crash recovery, no sibling impact | ➖ Not applicable |
| Multi-agent coordination | ✅ PM coordinator, kanban task graph, spawned subagents with persona and model config | ❌ Not supported |
| Shell execution | ✅ Stateful PTY with CommandPolicy AST filtering + approval gating | ➖ Not applicable |
| Browser automation | ✅ Native Playwright integration | ➖ Not applicable |
| Adaptive learning | ✅ NxCritic LSTM reward model on execution trajectories — compounds across sessions | ❌ None |
| Planning phase gating | ✅ Read-only Synapse tools during planning; write tools unlocked only in execution phase | ➖ Not applicable |
| Degraded mode fallback | ✅ Transparent grep/file_glob fallback when Synapse unavailable — agents unaffected | ➖ Not applicable |
| Parallel decision evaluation | 🟡 Concurrent tool dispatch; sequential planner rounds | ✅ All questions evaluated in single forward pass |
| Latency (per decision) | Seconds–minutes per planner round | ✅ 70–500ms per call |
| Cost (high-volume classification) | ❌ Full frontier LLM token costs | ✅ $0.042/MTok — near-zero at scale |
| Local-first / BYOK | ✅ Fully local, no cloud dependency, BYOK multi-model routing | ❌ Cloud API only — US West Coast, data sent to TypeSafe servers |
| Data sovereignty | ✅ All data stays on developer's machine | ❌ Data processed by TypeSafe infrastructure |
| Model agnosticism | ✅ Any Anthropic/OpenAI/compatible model — configurable per subagent | ❌ Jev model only |
4. Deep Capability Analysis
4.1 Context Management
Myelix winsTypeSafe: Assumes the caller provides correct, relevant, complete context. Their docs advise "include only the context relevant to the current questions" but provide no mechanism to determine what that is. Context assembly is the developer's responsibility entirely.
Myelix Agents: Context assembly is a first-class product feature:
- Synapse structural graph — semantically and structurally grounded context from live AST analysis, not raw text search
- 3-tier prompt cache — Stable (role, protocol, schema) → Context (intent, evidence, memory) → Volatile (current task, results) with 85–91% cache hit on stable tiers
- VWO overlay — child-agent worktrees pass their own
repo_idto Synapse; results correctly shadow parent chunks - Relevance-aware pruning — oldest journal entries pruned first; protected slots (intent_anchors, gate_violations, steer_directives) never pruned regardless of budget pressure
- Planning phase gating — Synapse tools are read-only during the planning phase; the capability catalogue unlocks write tools only in the execution phase, preventing premature mutations
4.2 Hallucination Prevention
Different layersTypeSafe: Schema enforcement at the model output layer. The output space is closed by the caller's schema. The model mathematically cannot generate a value outside the schema. Wrong-but-schema-valid answers are still possible if input context is incorrect.
Myelix Agents: Multi-layer prevention across the full agent lifecycle:
- Pre-write Synapse simulation — structural validation of proposed changes before any byte hits disk
- Action schema IR validation — planner action arrays validated against canonical action contracts before dispatch; malformed actions rejected before tool execution
- Five satisfaction gates — rule adherence → plan adherence → test coverage → diff safety → final satisfaction; each gate produces structured violation data, not generic failures
- CommandPolicy AST filtering — shell commands parsed by Elixir AST before execution; not regex, not keyword match — parse tree analysis
- Per-action approval gating — configurable approval requirement per tool, with allowlist persistence for trusted patterns
- Autonomy tiers — Tier 0 (full autonomy) through Tier 3 (approve every action) configurable per subagent
TypeSafe prevents schema hallucinations in individual typed decisions. Myelix Agents prevents systemic errors across multi-step stateful workflows. These are different layers of the same problem.
4.3 Confidence and Uncertainty
TypeSafe winsTypeSafe: Native calibrated probabilities from RLCD training. Mathematically derived confidence score from the probability distribution (not model-generated). Three-zone routing pattern documented: high confidence → act autonomously, medium → human review, low → escalate. The probability array is accessible, not just the top answer.
Myelix Agents: Confidence signals exist but are less precise:
- Gate checks are structured pass/fail — not probabilistic
- Autonomy tiers control risk tolerance at configuration time, not per-decision
- NxCritic assigns trajectory-level scores, not per-action probabilities
- The planner may express uncertainty in natural language but this is not programmatically queryable
This is the most significant gap. Per-decision calibrated probability is TypeSafe's core innovation and Myelix Agents has no equivalent. This is also the primary integration opportunity.
4.4 Speed and Cost for Decisions
TypeSafe wins by orders of magnitude| Metric | Myelix Agents | TypeSafe Jev |
|---|---|---|
| Latency per decision | Seconds–minutes per planner round | 70–500ms |
| Input cost | $0.20–$10/MTok (frontier LLM rates) | $0.042/MTok |
| Output cost | ~5x input cost typical | Free (too cheap to meter) |
| Sampling method | Sequential autoregressive | Parallel (single forward pass) |
Myelix Agents is not designed for high-volume real-time classification workloads. TypeSafe is exactly that. This is not a weakness — it is product scope.
4.5 Agent Capability and Stateful Workflows
Myelix wins — no contestTypeSafe: Deliberately stateless. No agent loops. No memory. No multi-step planning. Their philosophy: "code owns the workflow, model handles narrow judgements."
Myelix Agents — current capabilities:
- Planner loop with round budget and gate protection
- PM coordinator for cross-session orchestration and kanban task decomposition
- Subagent spawning with configurable persona, model, tool access, and autonomy tier
- Steer directives — developer corrections injected mid-run as protected prompt slots without interrupting execution
- Evidence ledger — structured facts accumulated across rounds, never pruned
- Clarification log — ask_user answers persisted as LearningEpisodes and fed back via MemoryContextProvider
- Session learnings and anti-pattern signatures stored for future session improvement
- Dream path trajectory data extracted for NxCritic training
- Current plan store — plan approval required before execution, stored separately from conversation
- VWO (Virtual Worktree Overlay) — child agents in worktrees see correctly scoped Synapse results
TypeSafe deliberately does not compete here. They solve a different problem.
4.6 Code Intelligence
Myelix wins — TypeSafe has noneTypeSafe: No code understanding. Text input only. Callers provide code state as raw text; the model has no structural awareness.
Myelix Agents via Synapse:
- Persistent structural AST knowledge graph — 50+ languages, local, cross-session
- Caller/callee graph traversal with transitive depth
- Change blast-radius analysis — transitive dependency radius, test coverage gaps, contract-sensitive signals
- Edit readiness assessment — relationship graph + linked tests + dependent files before any write
- Execution path tracing — ranked paths through dispatch, branching, side-effects
- Staged indexing — quick_pass for discovery, deep_analysis for full AST, debounced on file change
- BM25F keyword search + sqlite_vec KNN semantic search in a single unified call
- Read-only mode during planning phase via capability catalogue phase gating
- Degraded fallback inventory — transparent grep/file_glob when Synapse is unavailable
5. SWOT Analysis
Myelix Agents Strengths (relative to TypeSafe as reference)
| Strength | Evidence from Live Codebase |
|---|---|
| Full multi-agent orchestration | Planner loop, PM coordinator, subagent spawning, kanban, steer directives, plan approval store |
| Structural code intelligence | Synapse AST graph — callers, callees, tests, simulation, VWO overlay, planning phase gating |
| Persistent cross-session memory | Evidence ledger, clarification log (LearningEpisodes), session learnings, NxCritic trajectory store |
| Multi-layer hallucination control | Synapse pre-write simulation, action schema IR, five gates, CommandPolicy AST, autonomy tiers |
| BEAM fault tolerance | OTP supervision trees, microsecond restart, no sibling impact, preemptive scheduling |
| Data sovereignty + local-first | Fully local BEAM runtime, BYOK, no cloud dependency, Synapse runs on-device |
| Adaptive learning | NxCritic LSTM on dream path trajectory data — compounds, each session improves future sessions |
| Degraded-mode resilience | Capability catalogue fallback inventory — transparent grep/file_glob, agents unaffected by Synapse unavailability |
Myelix Agents Weaknesses (where TypeSafe leads)
| Weakness | Impact |
|---|---|
| No calibrated per-decision probability | Cannot build confidence-gated routing or uncertainty escalation natively |
| Latency (seconds vs milliseconds) | Unsuitable for real-time application paths requiring <100ms response |
| Cost for high-volume classification | Frontier LLM token costs prohibitive at millions of decisions per day |
| Sequential planner rounds | Independent questions still resolved across successive rounds, not in a single parallel forward pass |
6. Integration Opportunities
These are not speculative — they are concrete integration points that could be built today using the Myelix Agents tool contract system:
INTEGRATION 01 — HIGH VALUE
Jev as a Cortex Tool for Confidence-Gated Gate Evaluation
Register a
jev_noul tool in the Myelix Agents tool registry. At gate evaluation time, instead of a binary LLM pass/fail, run a Jev Noul question against Synapse-provided structural state: "Is this diff safe to merge given the impact radius?" at 70ms and $0.042/MTok. The probability output maps directly to the autonomy tier threshold — confidence >0.90 auto-approves; 0.60–0.90 requests human confirmation; <0.60 blocks.INTEGRATION 02 — HIGH VALUE
Intent Routing via Jev Choice
The PM coordinator currently routes incoming requests using a planner LLM call. Replacing this with a Jev Choice call — options:
code_change, research, refactor, bug_fix, test_write, documentation — would reduce routing latency from seconds to <100ms and reduce routing cost by ~97% at high volume. The full probability distribution enables confidence-aware routing (low-confidence requests escalated to human triage).INTEGRATION 03 — MEDIUM VALUE
Diff Risk Scoring via Jev Score
Before the agent writes a proposed change, pass a structured representation of the Synapse blast-radius output to Jev as a Score question: "How risky is this change on a scale of 1–5?" A risk score of 4–5 triggers elevated autonomy tier (more gate checks, human approval required); score 1–2 allows full autonomy. 70ms, near-zero cost.
INTEGRATION 04 — MEDIUM VALUE
Train a Cortex-Domain Decision Model
NxCritic's dream path learning already collects labelled trajectory data at scale. This data could bootstrap a Cortex-domain decision model using RLCD-inspired training — effectively building a Myelix-specific Jev for hot-path gate evaluation, intent classification, and risk scoring without sending data to external infrastructure.
7. Strategic Verdict
VERDICT — NOT COMPETITORS
Different layers. Complementary primitives.
TypeSafe builds a decision primitive — a typed, probabilistic function call at 70ms. Myelix Agents builds the orchestration runtime that constructs the context those decisions need, executes their consequences safely across multiple steps, verifies the results through five gates, and learns from the trajectory. They sell to different buyers at different budget scales and solve fundamentally different problems.
THREAT ASSESSMENT — LOW TO MEDIUM
TypeSafe could become a threat if they expand upward.
If TypeSafe adds memory, context management, and agent loops, they encroach on the orchestration layer. Their team quality (Diogo Almeida built the RLHF training methodology at OpenAI) makes upward expansion credible. Current threat: low. If they announce workflow primitives or stateful sessions: monitor closely and accelerate the Jev integration to create switching costs.
Recommended Positioning
"TypeSafe gives you fast, typed decisions at the leaf nodes of your software. Myelix Agents gives you the autonomous engine that constructs the context those decisions need, executes their consequences safely across a verified multi-step workflow, and compounds intelligence session after session. Neither product makes sense without the problem the other solves."
A Myelix Agents × TypeSafe integration — Jev for <100ms decision routing and gate evaluation, Myelix Agents for stateful orchestration and code intelligence — tells a compelling enterprise automation story that neither company can tell alone.