FutureAGI

@FutureAGI list all agent definitions in our workspace, filtering for production assistants that are currently active

Found 12 agent definitions matching your criteria. The production assistants include: CustomerSupportBot (assistant_id: asst_abc123, type: conversational, last modified 3 days ago), SalesQualifier (assistant_id: asst_def456, type: lead-scoring, last modified 1 week ago), and OnboardingGuide (assistant_id: asst_ghi789, type: workflow, last modified 2 weeks ago). Each definition includes configuration for model selection, system prompts, and tool bindings. A next_cursor is available to retrieve the remaining 9 definitions in subsequent requests.

This example demonstrates FutureAGI's agent registry capabilities. The MCP returns metadata about configured agents but does not execute them or access their runtime state. Useful for auditing which assistants exist in your workspace before running evaluations or tests against them.

@FutureAGI show me all annotation queues with pending items, I need to prioritize which datasets need human review this week

Retrieved 4 annotation queues with pending work. 'Customer Intent Labeling' has 847 pending items (status: active, created 12 days ago), 'Safety Classification' has 203 pending items (status: active, created 5 days ago), 'Sentiment Tagging' has 56 pending items (status: paused, created 18 days ago), and 'Entity Extraction QA' has 1,402 pending items (status: active, created 22 days ago). The Entity Extraction queue has the largest backlog and may require additional annotator capacity.

This read-heavy example showcases FutureAGI's annotation workflow visibility. The MCP surfaces queue metadata and item counts but cannot modify queue status or assign annotators. Teams using FutureAGI for human-in-the-loop labeling can quickly identify bottlenecks without opening the web UI.

@FutureAGI pull all evaluation tasks for project 'Q1-ChatbotRefresh' and summarize which tests are failing most often across our last 50 runs

Found 8 evaluation tasks in the Q1-ChatbotRefresh project. Analyzing recent test results: 'Greeting Accuracy' passes 94% of runs, 'Multi-turn Context' passes 78% (failing on long conversations with topic switches), 'Refund Policy Compliance' passes 61% (frequently misquotes return windows), and 'Escalation Trigger' passes 88%. The Refund Policy Compliance task shows the steepest decline over the past two weeks, suggesting recent prompt changes may have introduced regressions. Consider reviewing the system prompt updates from March 8th.

This synthesis example pairs FutureAGI's eval task metadata with AI reasoning to surface patterns. The MCP lists tasks and their configurations but does not execute tests or return granular run logs—the AI infers trends from task names and hypothetical pass rates. Requires read access to the specified project.

When you need eval rigor for customer-facing AI agents

A 6-person AI product team ships a support chatbot every two weeks. Before each deploy, they run regression tests against 200 annotated conversations to catch hallucinations or tone drift. FutureAGI's eval task and run test tools let them script these checks in Switchy—no context-switching to a separate platform. The annotation queue tools help the PM triage edge cases flagged by the model. This works when your test corpus is under 500 examples and you're running evals weekly or more. If you're doing one-off experiments or your dataset is still under 50 rows, the setup overhead outweighs the benefit—stick with ad-hoc prompting in a shared chat. If you're shipping agents to real users and need repeatable QA, this MCP closes the loop.

When your design team needs synthetic user feedback at scale

A 3-person UX team at a B2B SaaS startup wants to stress-test a new onboarding flow before user interviews. They load 12 personas (enterprise admin, solo freelancer, skeptical CFO) into FutureAGI and simulate 50 walkthroughs in an afternoon. The List Personas and List Agent Definitions tools let them browse and trigger simulations from Switchy without learning a new UI. The dataset tools let them export transcripts for thematic analysis. This pays off when you're iterating fast and can't wait for live user sessions—or when your ICP is hard to recruit. If your product is consumer-facing or you already have a steady stream of real user feedback, synthetic personas add noise more than signal. For early-stage B2B teams validating workflows before launch, this shortcut is worth the API key.

When you're building a labeled dataset for model improvement

A 4-person ML team is fine-tuning a domain-specific model for legal document summarization. They need 2,000 human-reviewed examples, but labeling is slow and inconsistent across contractors. FutureAGI's annotation queue and label tools let them route batches to reviewers, track progress, and enforce a shared taxonomy—all queryable from Switchy during standup. The dataset tools let the eng lead spot-check quality without opening a separate dashboard. This setup makes sense when you're managing 500+ annotations and multiple labelers. If you're under 100 examples or working solo, a spreadsheet is faster. If you're coordinating a distributed labeling workflow and need audit trails, this MCP keeps the process visible in your shared workspace.