Platform V2: Skills, Learning, 4 Team Modes, HITL, A2A, Evaluations — thinnestAI's Biggest Update Yet
The Biggest Release in thinnestAI History
When we started building thinnestAI, the vision was simple: make it possible for anyone to deploy production-grade AI agents without writing backend infrastructure. Today, we're taking the largest single step toward that vision with seven interconnected features that fundamentally change what your agents can do, how they learn, and how they work together.
This isn't a roadmap. Every feature described below is live, configurable, and ready to deploy — right now.
1. Agent Skills: Modular Capabilities You Build Once and Reuse Forever
Every agent you've ever built started from scratch. You wrote the same compliance instructions, the same qualification prompts, the same tone guidelines — over and over. Agent Skills end that cycle.
A Skill is a packaged bundle of instructions, tool configurations, and knowledge references that you create once and attach to any agent with a single click. Think of them as plug-and-play capability modules.
- Create a "BFSI Compliance" skill with regulatory disclaimers and restricted topics — attach it to every customer-facing agent in your organization
- Create a "Lead Qualification" skill with BANT questions, scoring logic, and CRM sync tool — drag it onto your sales agent
- Create a "Brand Voice" skill with tone instructions and example phrases — apply it globally
When you update a skill, every agent using it inherits the change instantly. No redeployment. No per-agent editing. One change, organization-wide impact.
Skills are prioritized — attach multiple skills and control the order in which their instructions are injected into the system prompt. Compliance first. Personality last. You decide.
2. Unified Learning: Agents That Get Smarter With Every Conversation
Your agents handle hundreds of conversations. Some go perfectly. Some don't. Until today, the lessons from those interactions disappeared the moment the session ended.
The Unified Learning System captures, indexes, and applies insights from every interaction:
- Automatic capture: When users give feedback (thumbs up/down, corrections, ratings), the system extracts a learning: "When handling refund requests, always ask for the order number first"
- Manual capture: Your team can add learnings directly — company policies, process changes, domain knowledge
- Semantic retrieval: Before every conversation, relevant learnings are pulled via vector search and injected as context — the agent remembers what worked and what didn't
- Organization-wide learnings: Capture insights that apply to every agent — brand guidelines, compliance rules, universal best practices
Each learning carries a confidence score, an application counter, and a source trace. You know exactly where every piece of institutional knowledge came from and how often it's being used. Invalidate outdated learnings with a click.
3. Four Team Execution Modes: The Right Orchestration for Every Workflow
Before this release, teams had one mode: the leader delegates. That works for some workflows, but real-world agent orchestration demands more flexibility. Now you have four modes:
- Coordinate: The leader delegates specific tasks to specific members based on expertise. Classic orchestration.
- Route: The leader examines each incoming request and sends it to the single best member. Only one agent runs per query — fast and cost-efficient.
- Collaborate: Every member works on every request in parallel. The leader synthesizes all outputs into a unified response. Maximum coverage, higher cost.
- Tasks: Each member has a fixed, pre-defined assignment. No dynamic delegation — each step executes its assigned work in sequence. Pipeline-grade reliability.
Select the mode from the Team Configuration modal's visual mode picker — four cards, one click. The underlying Agno framework handles the rest.
4. Human-in-the-Loop: Guardrails That Actually Guard
Autonomous agents are powerful. Unchecked autonomous agents are dangerous. HITL gives you surgical control over when and where humans need to intervene.
Four modes cover the full spectrum of oversight:
- Disabled: Fully autonomous (default)
- Blocking: Workflow pauses at designated gates until a human approves
- Confirm Tools: Agent operates freely but pauses before executing specific tools (e.g., "send_email", "create_invoice")
- Audit: Agent runs autonomously but every action is logged for post-hoc review
The Approvals dashboard shows all pending requests with priority badges, agent context, and one-click approve/reject. Auto-approval timers prevent bottlenecks for low-risk gates.
For regulated industries — BFSI, healthcare, legal — HITL is not optional. It's the feature that makes the difference between a proof-of-concept and a production deployment.
5. A2A Protocol: Your Agents Speak to the World
Google's Agent-to-Agent (A2A) protocol is an open standard for cross-platform agent interoperability. When you enable A2A on a thinnestAI agent, it publishes a machine-readable Agent Card at a well-known URL. Any A2A-compatible agent — on any platform — can discover and invoke your agent.
- One toggle: Enable A2A from the agent settings panel
- Instant publishing: Your agent card is live at
/.well-known/agent.json - Capability tags: Describe what your agent does (text-generation, data-lookup, task-execution) so others can find it
- Discovery API: Search for other A2A-enabled agents by capability or description
- Full audit log: Every incoming invocation is logged with caller, message, status, and timestamp
A2A transforms your agent from an isolated tool into a node in a global agent network. Sales agents invoke billing agents. Support agents invoke technical specialists. Partner agents invoke your services without sharing infrastructure.
6. Agent-as-Judge Evaluations: Quality You Can Measure
"Is my agent good?" is no longer a vibes-based assessment. The Evaluation system uses a dedicated LLM judge to score your agent's conversations against configurable criteria.
- Default criteria: Helpfulness (0–10), Accuracy (0–10), Tone (0–10), Instruction Following (pass/fail)
- Custom criteria: Add your own — "Did the agent attempt an upsell?", "Was the compliance disclaimer included?"
- Single or batch: Evaluate one conversation or thousands
- Score trends: Track quality over time with visual charts
- Judge reasoning: Every score comes with an explanation — you know why the agent scored low, not just that it did
Combined with the Learning System, evaluations create a closed loop: Evaluate → Identify Weakness → Capture Learning → Re-evaluate. Continuous improvement, automated.
7. Multi-Engine Scraping Hub: Four Scrapers, One Interface
Different websites require different scraping approaches. The Scraping Hub provides four engines behind a single, unified interface:
- Standard (BeautifulSoup): Fast, lightweight, reliable for static sites
- Crawl4AI: AI-powered extraction with smart content parsing
- Firecrawl: Cloud-based with full JavaScript rendering
- Scrapling: Stealth scraping with three fetcher tiers, proxy support, and anti-detection
Each engine has its own visual card with feature badges, unique accent colors, and engine-specific configuration. Advanced settings include CSS selectors, proxy support, stealth mode, and JS rendering toggles.
Content deduplication via SHA-256 hashing ensures re-scraping the same pages doesn't bloat your knowledge base. Real-time SSE progress shows extraction as it happens.
Also Shipping: 7 New Tools, 5 Knowledge Sources, If/Else Branching
Beyond the seven headline features, this release includes:
- 7 new agent tools: Scrapling, MCPTools, Seltz, Unsplash, Redshift, UserFeedback, DashPostgres
- 5 new knowledge sources: Azure Blob Storage, SharePoint, Private GitHub repos, Excel (.xls/.xlsx), AST-based code chunking
- If/Else branching: Conditional flow routing with keyword, regex, and LLM-evaluated conditions
- RBAC: Role-based access control with owner/admin/editor/viewer roles
Everything Works Together
The real power of this release is the interconnection:
- Skills + Learning: A skill teaches the agent what to do. Learnings teach it how to do it better.
- HITL + Teams: Coordinate mode with blocking gates — the leader delegates, but high-stakes steps require human sign-off.
- Evaluations + Learning: Low evaluation scores automatically suggest learnings to capture. The feedback loop closes itself.
- A2A + Teams: External agents can invoke your team, triggering the full orchestration flow — coordinate, route, collaborate, or pipeline.
- Scraping Hub + Evaluations: Scrape competitor documentation, feed it to your knowledge base, then evaluate whether your agent's responses are better.
Get Started
Every feature in this release is available on all plans, including the free tier. No feature gates. No enterprise-only restrictions. Build the most sophisticated agent workflows you can imagine — and deploy them today.
No credit card required • All features included • Deploy in minutes