Changelog

We’re rolling out Fuse, our new architecture for orchestrating multiple specialized agents in a single workflow.
With Fuse, you can now:

• Chain agents with task-specific prompts

• Enable parallel agent execution with shared memory

• Call external tools from within agent trees

The beta includes support for function calling and shared context across agents.

Fuse is available for Pro and Enterprise plans.

📦 New: GPU Auto-Scaling (for Inference APIs)

Our API now supports automatic GPU scaling based on real-time traffic.
This helps reduce cold starts and ensures low-latency inference even during usage spikes.

• Support added for NVIDIA A100, H100

• Billing adjusts dynamically based on load

• Requires no configuration — just deploy your model

🧠 Improved: Model Updates

• Upgraded our default CodeGen-7B endpoint to v2.1 — better accuracy, fewer hallucinations

• DocQA model now supports 150k token contexts

• Improved multi-language support in Chat endpoint (added Korean, Dutch, Polish)

🔐 API Changes

• New /v1/agents/run endpoint for orchestrated multi-agent flows

• Deprecated /v1/tasks/create — use /v1/agents/launch instead

• API keys can now be scoped per model, feature, or environment (dev/staging/prod)

🧪 Labs

• Internal tests running for speech-to-code pipeline (using Whisper + CodeT5)

• Early access to fine-tuned vision transformer (ViT-x3) for document parsing

• Testing memory-aware agents with local context retention beyond sessions

🛠 Fixes

• Fixed a memory leak in real-time embeddings endpoint

• Resolved an auth issue causing 401 errors on PUT /models/train

• Improved latency for European region (Frankfurt): -35ms avg per call

Cron Patterns

You can now assign a cron‑style schedule to any agent. Pick days, hours, or minute intervals directly in the panel, and preview the next five run times before saving. Schedules respect your workspace time zone, while the engine stores everything in UTC to avoid daylight‑saving surprises.

• Hourly, daily, weekly, or custom cron strings

• One‑click “every X minutes” presets

• Weekend‑only or business‑hours windows

• Pause/resume without losing the pattern

• Human‑readable next‑run preview

• Safe‑guard against overlapping executions

Conditional Triggers

Need more control? Add a conditional rule so the job fires only when data changes or KPIs cross a threshold. Combine multiple conditions with AND/OR logic, and see a live evaluation in the sidebar before enabling.

Smart intent routing

This month we introduced smart intent routing and an updated policy wizard. The router now inspects each prompt’s intent, automatically steering Q&A traffic to fast models while sending creative tasks to richer LLMs. A new preview pane shows token cost and latency projections for every rule before you hit Save, so you can tune policies with confidence.

Patch Notes

• Earlier in the year we focused on stability, observability, and fine‑grained control. See the highlights below.

• Added intent classifier for real‑time model selection

• New bulk‑edit mode for adjusting weights across many routes

• Latency budget slider to cap 95th‑percentile response times

• Per‑model retry limits to prevent runaway loops

• CSV export of routing logs with token counts and costs

• Alerting webhook for failover events and degraded endpoints

• UI fix: dropdown no longer resets on browser back‑navigation

• Security patch for rare race condition in auth middleware