What We Shipped in July 2026

June was about control after you ship. July is about what your agents run on. Connic Models puts 30 EU-hosted models from nine creators behind one set of IDs you name in a line of YAML, Project credit collapses billing into a single balance that covers everything a Project consumes, and the Marketplace is now the one catalog for templates, connectors, and retrieval sources. Underneath, the deploy gate learned to script human approvals and mock lifecycle code, model retries stopped replaying your agent, long sessions compress instead of overflowing, and a new AI Governance workspace turns EU AI Act readiness into a record you can export.

Connic Models: Managed EU Inference

The biggest release of the month. Connic Models unifies 30 models from nine creators behind one set of connic/* IDs that every Project can use out of the box. Name one in agent YAML and deploy: no provider account, no API key, no region flag. Every call runs on EU inference capacity, with request retries and cross-provider failover handled for you.

model: connic/glm-5.2

# or a latency-optimized execution profile:
# model: connic/kimi-k3-fast

There is no key to create, paste, or rotate: managed tokens bill straight from Project credit at published EUR rates per model, and BYOK stays available whenever you want your own provider. Models with a -fast ID route to latency-optimized capacity at the same price; outputs may differ from the standard ID, so treat a switch like a model change and test.

Read the Connic Models announcement or browse the model catalog with exact IDs and prices.

Project Credit: One Balance for Everything

Billing now runs on Project credit, one balance that covers everything a Project consumes: runs, compute, storage, synced retrieval items, and connic/* model tokens. Basic Projects start with a one-off €25 welcome credit and can keep topping up from there. A Developer or Pro subscription arrives as the same amount of monthly credit, so the subscription itself is credit to spend, and plans unlock features and higher limits as you grow. Enterprise Projects stay postpaid with a monthly invoice.

Standard Projects are prepaid by design: usage stops when the balance is insufficient, and capped auto-refill tops credit up before that happens, never past your cap. The full rate card lives in one place, the pricing page, and the usage docs show how consumption maps to credit.

The Connic Marketplace

The Marketplace launched at the start of the month as the single catalog of building blocks for Connic projects. Every listing is published by Connic and free to install, and the same catalog is browsable on the website and inside the dashboard.

connic init my-project --templates=invoice

Read the Marketplace announcement or browse the Marketplace directly.

Testing: Scripted Approvals and Lifecycle Mocks

The deploy gate can now test flows that pause for a human. approval_decisions scripts the outcome for a pending tool approval, so a case can assert what happens when a refund is approved, rejected, or times out, without a person clicking through. The same mocks module that already replaces custom tools can now replace middleware phases, tool hooks, and custom guardrails. And results stream in live: each test case reports as it finishes instead of after the whole suite.

tests:
  - name: approves_refund_then_rejects_notification
    builder: create_charge_then_refund
    approval_decisions:
      - tool: billing.refund
        params: params.charge_id == context.charge_id
        decision: approve
      - tool: notifications.send
        decision: reject
        reason: Do not send notifications in this scenario

• •Approve, reject, or timeout: Every branch of an approval flow becomes a testable case, matched by tool and optionally by parameters
• •Strict lifecycle flags: strict_hook_mocks, strict_middleware_mocks, and strict_guardrail_mocks each fail a case before an unmocked real phase executes, so a missing mock fails loudly instead of running real code
• •GitLab, fully covered: Self-managed GitLab instances connect with a personal access token, and merge request testing reports connic/pr-tests before merge

See the test YAML docs and the fixtures and mocks reference for the full rules.

A/B Testing: Confidence Mode

A/B tests now come in two shapes, and you pick the one that matches the question you are asking.

Read the A/B testing docs for both modes.

Retries That Do Not Replay Your Agent

When a provider returns a transient error mid-run, Connic now retries exactly that model request. Tools that already ran do not run again and middleware does not replay, so a flaky provider minute cannot trigger duplicate side effects. A configured fallback model takes over at the same call boundary, and the new initial_delay option tunes the first backoff interval alongside the existing attempt and delay caps.

retry_options:
  attempts: 3
  initial_delay: 5
  max_delay: 30

The trace stays honest, too: each retried attempt shows up as a sibling span with its backoff metadata, and provider error messages arrive intact instead of flattened into a generic string. See runtime controls for the full retry reference.

Context Compression for Long Sessions

Long-running agents no longer die at the context window. When a provider context-window error hits, the runner compresses older history and oversized tool results, then retries the call. keep_recent_messages keeps the newest turns verbatim, and an optional max_prompt_tokens budget compresses proactively before a later call instead of waiting for an error.

context_compression:
  enabled: true
  model: connic/glm-5.2
  keep_recent_messages: 12
  session_history:
    interval: 4
    keep_recent_runs: 1

Between runs, session_history compacts stored history on an interval while keeping the most recent runs untouched. Both paths can use a smaller dedicated model for summaries while normal calls stay on the agent model, and every compression appears as its own trace span with token counts. Read the deep dive on context compression for how the pipeline works.

Per-Call Tracing and Guardrail Outcomes

Every model call in a run now gets its own trace span with token usage attached: input, output, thinking, and cached input, shown as a compact pill on each span in run details. Long agent loops stop being one opaque number and become a sequence of calls you can inspect individually.

Guardrail interventions are now stored on the run itself and appear in the runs table as filterable outcomes. Finding every blocked injection attempt is now one filter, not a scroll through run details.

Automatic Bridge Routes

Some protocols connect to one service and get handed a different hostname for the next connection: Redis Sentinel returning the current master, Kafka metadata, database failover clients. Until now that handoff needed the bridge suffix on the returned destination, a name you often cannot change. Automatic routes fix this: declare exact hostnames or a safe regex per bridge, and matched traffic from your code routes through the bridge with the ordinary hostnames unchanged.

{
  "routes": [
    {
      "match_type": "exact",
      "target": "redis-sentinel.internal",
      "port": 26379
    },
    {
      "match_type": "regex",
      "target": "^redis-[a-z0-9-]+\\.internal$",
      "port": 6379
    }
  ]
}

Routes are managed under Project Settings in the Bridge section, or through the routes API. Matching stays predictable: explicit bridge hostnames take priority, then exact routes, then regex routes, and regex targets accept only anchored, narrow patterns. The bridge docs include a complete Redis Sentinel walkthrough.

AI Governance for the EU AI Act

Enterprise Projects get a new AI Governance workspace that turns EU AI Act readiness into a working record next to your agents, runs, and deployments. Register AI systems, record preliminary provider and deployer assessments, then generate controls from those assessments and work them to completion. It is a documentation and evidence workflow, not a legal determination: final classification stays with you and your counsel.

• •Scoped evidence: A snapshot captures one system, selected environments, and a time window, exportable as an immutable, hash-verified archive for auditors and counsel
• •Article 50 in one place: Document how disclosure obligations are met, including upstream provider attestations
• •Incidents included: Log and track incidents so the evidence trail reflects what actually happened

Want this feature? Contact your account manager to get access. The AI Governance docs cover the full workflow.

More Improvements

• •Retrieval: The Knowledge Base is now Retrieval across the product, with existing config keys and tool names still working, and namespaces can be deleted straight from the dashboard, child namespaces included
• •A home for dashboards: Create, reorder, and delete dashboards on a dedicated page, and pin up to 15 to the sidebar
• •Selective drains: Log and audit drains take a filter_expression, validated as you type, and forward only matching entries
• •Markdown tables: Run details render tables in agent output, useful for report-style agents
• •Bulk judge triggering: Point a judge at a date range, preview how many runs it will score, and trigger the whole batch from the judge page
• •Sensitive variables in the raw editor: Prefix a line with ! in the raw environment editor, as in !API_KEY=value, to mark it sensitive
• •Structured trigger payloads: trigger_agent and trigger_agent_at accept dicts and lists natively, so orchestration code passes structured JSON without stringifying
• •Docs, restructured: The documentation is reorganized into Build, Test, Platform, Connectors, and Reference