Magic Evolve · Continuous auto-improvement

Every conversation makes the next one better.

Name: Magic Evolve
Brand: Magic Sites
Price: 19 USD
Availability: InStock

Memory consolidation, preference learning, skill extraction, workflow autopsy, A/B promotion. Magic Evolve runs the eval loop on idle compute so your agents compound while you sleep.

Reference architecture from public literature -- STaR, Voyager, Reflexion, Self-Refine, DPO, RLAIF. Built on edge AI. Bundled with every Pro tier across the Magic stack.

Start Evolving for $19/mo Compare tiers Edge AI compute · no per-token markup

The eval loop

Four loops, one durable improvement engine.

Each loop is a primitive from public literature, productized as a multi-tenant managed workflow. They activate by tier; they all share the same eval harness, ground-truth replay, and ledger-based promotion router.

01

Memory consolidation

All tiers

Reflexion-pattern episodic-memory consolidation. Last seven days of conversation distilled into a system-prompt addendum. Frees the context window without losing what worked.

02

DPO preference learning

Pro and up

Every thumbs-up, every thumbs-down, every accept-or-edit becomes a Rafailov-style preference pair. Better answers without you writing prompt rules.

03

Skill extraction

Scale and up

Voyager-pattern skill library. When your agent runs the same tool sequence three times, Magic Evolve proposes a reusable skill and adds it to your toolbox.

04

Workflow autopsy

Scale and up

Reflexion-pattern post-mortem on failed workflow runs. Replay the failure, propose a patch, run the patch against the original input, ship if it passes.

Pricing

Four tiers. Same eval-loop core. Bigger budget at every step.

Lowest tier is paid -- per Magic Sites policy, we never bake a free tier into a SKU with hard compute cost. Founding-rate credits available at signup.

Starter

Memory consolidation only. For founders who want their agent to remember what worked yesterday.

$19 /mo

✓ Memory consolidation loop
✓ 1M tokens / month (fast open-source models)
✓ Auto-applied weekly distillation
✓ Token-bucket dashboard
✓ Per-feature opt-in toggles
✓ 1 active improvement at a time
✓ Community support

Start with Starter

Pro

Scale

Skill extraction + workflow autopsy + custom evals + custom fine-tunes. For agents that compound monthly.

$249 /mo

✓ All Pro loops + skill extract + autopsy
✓ 100M tokens / month (open-source + Kimi)
✓ Customer-defined eval criteria (sandboxed code)
✓ Canary traffic split routing
✓ Custom fine-tune dispatch
✓ 25 concurrent improvements
✓ Priority support

Start with Scale

Enterprise

Dedicated improvement compute, white-label, on-prem deploy, custom training schedule, priority queue.

Custom

✓ All Scale loops + custom evaluator code
✓ Custom monthly token budget
✓ Dedicated routing + reserved compute
✓ On-prem deploy option
✓ Custom training schedule
✓ Custom concurrency
✓ Dedicated solutions engineer

Talk to sales

How it works

Three steps from signal to shipped improvement.

01

Capture signal

Conversations, tool traces, workflow runs, and feedback events stream into the per-tenant evolve queue. Your data, your tenant, never crossed.

02

Run the eval loop on idle compute

Off-peak windows trigger the self-improve consumer. Edge AI generates proposals; the eval harness scores them against frozen ground truth. Keep or discard, binary.

03

A/B promote, rollback in one row

Winners route through smart model routing with versioned routes and optional canary split. Rollback is a single ledger update -- no DNS, no redeploy.

Customer signal sources
   Magic Agent      Magic Workshop      Magic Loops      External (BYO API)
        |                  |                  |                |
        +------------------+------------------+----------------+
                                  |
                                  v
                  +---------------+---------------+
                  |  Job classifier worker        |
                  |  reads tier + token bucket    |
                  +---------------+---------------+
                                  |
   +----------+-----------+-------+--------+----------+----------+
   v          v           v               v          v          v
memory    distill       dpo            skill     autopsy    finetune
consol.   (Pro+)        (Pro+)         (Scale+)  (Scale+)   (Scale+)
   |          |           |               |          |          |
   +----------+-----------+---------------+----------+----------+
                                  |
                                  v
              Self-improve consumer (automation engine)
              step.sleepUntil(off_peak)  ->  budget gate
                                  |
                                  v
              Edge AI: Gemma 4 / Kimi K2.6 / fine-tune
                                  |
                                  v
                  Eval agent (per tenant)
                  keep / discard binary
                  improvement_proposals queue
                                  |
                                  v
              Smart promotion router
              versioned routes per tenant
              canary split (Pro+) -- rollback in one row

Across the Magic stack

One eval loop. Every agent gets better.

Magic Evolve runs on top of Magic Agent, Magic Workshop, and Magic Loops out of the box. Bring your own external agent and integrate via the standalone API.

Magic Agent: quality compounds

Every conversation feeds memory consolidation; every thumbs-up feeds DPO. Your agent gets smarter with traffic, not in spite of it.

Magic Workshop: skills extract themselves

Your workspace session writes a tool-call ledger to /workspace/.evolve/. After three repeats, a script lands in /workspace/.skills/ ready to call.

Magic Loops: workflows auto-optimize

Failed runs trigger autopsy. High-frequency successes trigger latency optimization. Both surface as accept-or-reject proposals in your admin.

BYO external agent: standalone API

Running Claude, Codex, Hermes, or your own stack? POST conversations and feedback to the Magic Evolve API; GET back proposals and the active version pointer.

Standalone tier

BYO agent? Same loop, different door.

Running Claude, Codex, Hermes, or your own stack? Subscribe to Magic Evolve standalone. POST conversation logs, thumbs-up / thumbs-down, tool traces, and workflow records. GET back proposals and the active version pointer. Route traffic through your own infrastructure. The platform owns the eval-loop core; you own the agent.

Start standalone Pro Talk to sales

SDK: @magic-sites/evolve-sdk · REST base: https://evolve.magicsites.website/api/v1/{tenant}/

Bundled across the Magic stack

Already on a Pro plan? You already have Magic Evolve.

Magic Sites Studio, Magic Agent Pro, Magic Workshop Pro, and Magic Loops Pro each include one Magic Evolve Starter. Magic Sites Agency, Magic Agent Scale, Magic Workshop Team, and Magic Loops Scale each include one Magic Evolve Pro. No extra subscription. Usage above the bundled tier upgrades to a paid Evolve line.

See Magic Sites pricing Magic Agent

How we compare

Productized eval-loop. Edge-native compute. No model lock-in.

Option	Price	Loops	Compute
Magic Evolve Pro	$79/mo	Memory + DPO + distillation + A/B	Edge AI (no markup)
Custom DPO pipeline	$2-10k once	You build it	You run it
LangSmith Plus	$39/seat	Eval + datasets	BYO models, BYO promotion
Helicone Growth	$80/mo	Logging + experiments	BYO promotion
OpenAI fine-tune flow	$8/M tokens	Fine-tune only	OpenAI lock-in

FAQ

Questions before you ship.

What does Magic Evolve actually do that prompt-engineering does not?

Continuously. Prompt-engineering happens once a week when a human notices something is off. Magic Evolve runs every off-peak window: distills what worked, scores what did not, proposes a fix, A/B tests it, promotes the winner. The eval loop is automated end to end.

Where does the reference architecture come from?

Public literature only. STaR + V-STaR for rationale fine-tuning, Voyager for skill libraries, Reflexion for episodic memory, Self-Refine for in-context iteration, DPO for preference optimization, RLAIF for AI-graded preferences. arXiv IDs in the design doc.

What models does it use?

Fast open-source models (Gemma 4) plus frontier-grade Kimi K2.6 routed through smart model routing. No OpenAI, no Anthropic. When custom fine-tuning ships GA, Scale tier activates fine-tune dispatch with no architecture change.

Is my data shared across tenants?

Never default-on. Every loop runs strictly on your tenant data with your token budget. Cross-tenant aggregation is a permanent opt-in for explicit research consent only -- it is not a default at any tier.

What gets surfaced for human review and what auto-applies?

Memory consolidation auto-applies on Starter (improves your agent on your data, no privacy surface). Pro and Scale customers see proposals in the admin queue with accept, reject, and replay-before-promote. Cross-tenant flows are never auto-applied.

How do bundled Evolve subscriptions work?

Magic Sites Studio, Magic Agent Pro, Magic Workshop Pro, and Magic Loops Pro each include one Magic Evolve Starter at no extra charge. Magic Sites Agency, Magic Agent Scale, Magic Workshop Team, and Magic Loops Scale each include one Magic Evolve Pro. Usage above the bundled tier upgrades to a paid Evolve line.

Can I bring my own agent?

Yes. The standalone tier wraps the same eval-loop core with a REST plus SDK surface. POST conversations, feedback, and workflow records; GET improvement proposals and the active version pointer. Your traffic stays on your infrastructure.

Can I cancel?

Anytime, in-app. Export your evolve_jobs, eval_runs, improvement_proposals, and promotion_ledger as JSON. No retention lock-in.

Stop prompt-engineering on Mondays. Start shipping a smarter agent every week.

Founding-rate Starter, Pro, and Scale credits at signup. Cancel anytime; export everything as JSON. No retention lock-in.

Start for $19/mo Talk to sales