What We Shipped in May 2026

May was about confidence before shipping. A new testing framework lets teams write and run tests for agents, deploy gates hold a deployment until those tests pass, and pull-request testing catches regressions before they merge. We also went deeper on visibility, tracing triggered and child runs and reporting on token usage and budgets, shipped custom domains for connectors, and added per-agent reasoning effort with cascading project defaults.

The Agent Testing Framework

Agents are now testable like any other code. Define cases in a tests/ file, give each one an input and a set of assertions, and connic runs them against the agents and records the results.

tests:
  - name: refunds_small_orders
    payload: '{"message": "Refund order A123, $40"}'
    expected_result: output.status == "refunded"
    expected_tool_calls:
      - process_refund

  - name: escalates_large_refunds
    payload: '{"message": "Refund order B456, $500"}'
    expected_result: output.status == "pending_approval"
    expected_child_agents:        # assert on a triggered child agent
      manager-approval:
        expected_triggered: 1

• •Run them anywhere: Kick off ad-hoc test runs from the dashboard, or test interactively during development with connic dev
• •Assert on triggered agents: Assertions can target both the child agents a run triggers and the top-level response
• •Coverage reports: See which agents and tools the tests actually exercise
• •Run a subset: Use connic test --filter to run only the cases that match

The full walkthrough is in A Testing Framework for AI Agents.

Deploy Gates and Pull-Request Testing

Tests are only useful if something acts on them. With deploy gates, a deployment is held until its test suite passes, configured per environment so staging and production can be gated independently.

connic also runs the tests on every pull request and reports the result back as a status check, with a link straight through to the deployment and its run traces.

connic/pr-tests   All tests passed   ✓   Details

24 passed, 0 failed.

• •Per-environment control: Turn gating on or off per environment, and point its tests at a separate test environment when needed
• •PR status checks: A single connic/pr-tests check shows pass or fail on the pull request, linking through to the deployment behind it
• •Queued ahead of execution: Deploy-gate runs are queued before the deployment starts, so a bad build never reaches the environment

Tracing Triggered and Child Runs

When one agent triggers another, the whole chain is now visible. Open a child run in a drawer stacked over its parent, step back up the chain, and jump between a trigger and the run it started, all without losing the current place in the run history.

• •Stacked drawers: Child runs open over their parent for navigation down and back up in one view
• •Linked navigation: Every triggered run records the run that started it, with inline links in both directions
• •Trigger source: A new trigger source on every run shows at a glance whether it came from a connector, another agent, a manual run, or a deploy gate

Usage and Budget Dashboards

The usage dashboard now reports token consumption and spend per agent alongside budget reporting, showing where cost is going and which agents are driving it.

Judges gained filter expressions for picking exactly which runs to evaluate:

context.tier == 'enterprise' and output.priority >= 3

• •Token and budget metrics: Per-agent token usage and spend, with agents that have no usage excluded from budget reports
• •Input widgets: Add dropdowns and inputs to a dashboard to filter every widget on it at once
• •Companion metrics: Show a secondary comparison value next to a widget's primary number for quick context

Custom Domains for Connectors

Connector URLs no longer have to live on a connic-hosted address. Point a controlled domain at the connectors and connic serves their webhook and websocket endpoints from it, so partner-facing URLs match the organization's brand.

# Before: connic-hosted
https://connect.connic.co/webhook/3b0a9d…

# After: custom domain
https://hooks.acme.com/webhook/3b0a9d…

• •Custom domain: Serve connector endpoints from an organization-controlled subdomain
• •Automatic verification: Add a domain and connic verifies DNS and issues a certificate, with the status shown as it progresses
• •Per connector: Point individual connectors at the domain that fits

Reasoning Effort and Cascading Defaults

Two additions to agent configuration. Set reasoning_effort per agent to trade latency for depth, and drop a _defaults.yaml in the agents/ folder to set common configuration once.

# Inherited by every agent in the project
reasoning_effort: medium
timeout: 120

Every agent inherits the defaults and overrides only what it needs, so a research agent can ask for reasoning_effort: high without repeating the rest of the project's settings.

More Improvements

• •Reusable connections: Define a connection once for centralized credential management and reference it from any agent or connector instead of repeating secrets
• •Conditional upserts: The new db_upsert tool inserts or updates a document in one call, and retrieval queries now support metadata filtering
• •Authenticated connectors: A new require_auth option locks down inbound webhook and websocket connectors
• •MCP header context: Interpolate run context into MCP server headers for per-request auth and routing
• •Clearer timeouts: More detailed timeout and cancellation messages, so a stopped run tells the exact reason it stopped
• •Redirect following: The web_read_page tool can now follow redirects through to the final URL