What We Shipped in February 2026

February was a packed month. Agents now get a managed database for persistent storage. The new templates library helps you get started faster, judges automatically evaluate agent output, and the Telegram connector opens up a new channel. We also shipped web page reading, persistent sessions, and several configuration improvements.

Agent Database

Every environment now comes with a managed database. Agents can store, query, update, and delete records in collections without any setup, migrations, or external services. Collections are created automatically on the first insert, and data is isolated per environment.

Six built-in tools give agents full control:

• ✓db_find: Query documents with filters, sorting, pagination, and field selection
• ✓db_insert: Store one or many documents at once
• ✓db_update: Update matching documents by filter
• ✓db_delete, db_count, db_list_collections

Add the tools you need to your agent:

name: order-processor
model: gemini/gemini-2.5-flash
system_prompt: |
  You process incoming orders and track their status.

tools:
  - db_insert
  - db_find
  - db_update
  - db_count

Filters support operators like $gt, $in, $regex, $and/ $or, and more. Every document gets automatic _id, _created_at, and _updated_at fields.

You can browse collections, inspect documents, view inferred schemas, and insert data manually from the new Storage > Database section in the dashboard.

Agent Templates Library

Starting a new agent project no longer means starting from scratch. The templates library offers pre-built agent projects that you can browse in the dashboard or install directly from the CLI.

Each template is a complete project with agents, tools, middleware, and a README. Available templates include:

• ✓Invoice processing with extraction, validation, and structured output
• ✓Customer support with ticket triage and RAG-powered responses
• ✓Telegram personal assistant with web search, notes, and persistent sessions
• ✓Stripe dunning for automated payment recovery workflows
• ✓And more: email helpdesk, fraud detection, document pipeline, compliance auditor, and others

Install templates when scaffolding a new project:

$ connic init my-project --templates=invoice
$ connic init my-project --templates=invoice,customer-support

Template dependencies and READMEs are automatically merged into your project. You can also browse all templates in the templates catalog in the dashboard, where each template shows its architecture, file structure, and code preview.

Judges

Judges automatically evaluate agent runs using LLMs. Define scoring criteria, set a sample rate, and let judges assess output quality on every run or a percentage of runs.

Each judge is configured with:

• •Scoring criteria: Named rubrics with descriptions and max scores (e.g. Accuracy out of 10, Tool Usage out of 5)
• •Trigger mode: Automatic on every completed run, or manual on demand
• •Sample rate: Evaluate 1-100% of runs to balance coverage and cost
• •Filters: Only evaluate runs matching specific conditions (status, agent name, context properties)
• •Notification threshold: Get alerted when average scores drop below a threshold

Evaluations produce per-criteria scores with reasoning, so you can understand exactly why a run scored the way it did. Create and manage judges from the new Judges tab in your project.

Telegram Connector

Connect your agents to Telegram. The new Telegram connector supports both inbound (receive messages, trigger agents) and outbound (send agent output as replies).

Setup takes a minute:

1. Create a bot via @BotFather and copy the token
2. Create a Telegram connector in Connic and paste the token
3. Link it to an agent

Your agent receives the message text, chat ID, and full message details. To reply, return JSON with the chat ID and response text:

{
  "chat_id": 987654321,
  "text": "Your order #12345 is on its way!"
}

Web Page Reading

The new web_read_page tool lets agents fetch any URL and receive the page content as clean markdown. This pairs well with web_search: search first to find relevant pages, then read them in full.

tools:
  - web_search
  - web_read_page

The agent passes a URL and gets back the rendered content as markdown, making it easy to extract information, summarize articles, or pull data from documentation.

Persistent Agent Sessions

Agents can now maintain conversation history across multiple requests. Sessions are keyed by a value from the trigger payload or middleware context, so each user or chat thread gets its own persistent memory.

session:
  key: context.chat_id    # from middleware
  ttl: 86400              # expire after 24h of inactivity

Session keys can reference middleware context (context.chat_id) or raw input fields (input.user_id). The optional TTL controls when inactive sessions expire. Sessions survive redeployments, and you can view and manage active sessions from Storage > Sessions in the dashboard.

More Improvements

• •Conditional tools: Control which tools are available per run using expressions over input and middleware context. Use a mapping instead of a plain string to attach a condition: web_search: context.search_enabled
• •Concurrency rules: Ensure only one run per key value (e.g. per customer or process ID) is active at a time. Choose queue to hold new runs until the active one finishes, or drop to discard duplicates
• •CLI rename: The connic dev command is now connic lint to better reflect what it does: validate your agent configuration locally before deploying