Synth

Synth is built for a specific kind of user: the AI engineer who treats agent development as a research discipline. If you're already comfortable with the CLI/SDK, Stack's live feed, artifacts, and handoffs create a tight loop between running an experiment and inspecting the evidence. The Sidecar monitor is a standout—it gives you a narrated stream of what your long-running agent is doing, letting you audit done-claims and steer the worker without sifting through every tool call. That's genuinely useful for sparse-reward tasks where you need to trust the agent's self-reports. Workshop v0.6.0 adds durability: optimizer workflows that survive restarts, local Qwen MLX SFT for fine-tuning without a GPU cloud, paired held-out evaluation to catch overfitting, and fail-closed evidence contracts that block claims without proof. The shift to a desktop app (v0.1 Friends Release) for Apple silicon makes the whole thing more approachable, though the first build is unnotarized and macOS 14/16GB minimums cut off older hardware. The cloud side—Managed Research—is where you scale. GELO and GEPA are hosted optimizers that run hill-climbs on your prompt or context automatically, and the GitHub Actions integration means you can wire prompt improvement into CI/CD. Open Research Factory publishes benchmark-owned grading, so you can show third-party proof that your agent improved. Where it falls short: this is not a tool for beginners. The learning curve is steep, the CLI is the primary interface, and there's no graphical fine-tuning UI. Pricing on paid tiers is opaque—Workshop Starter is $20/mo for 2,000¢ of usage, but Standard and Max are contact-sales with no public price. Hosted inference draws from your allowance, so heavy cloud usage can burn through credits quickly. If you're a researcher pushing on agent reliability, Synth gives you the receipts. If you need a turnkey eval platform with clear pricing, LangSmith or W&B will be smoother.