Agents that keep working on your busiest day.
No servers, queues or deploy pipeline to run.
connic builds, tests and runs every change you push. Requests that pile up wait their turn, a fallback model answers when a provider fails, and a release that fails its tests never replaces the working one.
Steady service when the events pile up.
Take a parcel service: scanners send events all day, business customers book pickups late at night, and upset recipients write three emails in four minutes.
Every scan starts the agent, but one parcel gets one run at a time. Scans that pile up meanwhile become one update with the latest data.
A pickup is booked at 23:14 just as the main model hits its limit. The fallback model writes the reply, and the booking stays as it is.
Emails from the same sender are handled one after another, so the second and third add to the case the first one opened. A person reviews the refund.
A release that fails its tests never replaces the working one.
connic builds every change, runs your tests and switches over only when all of them pass. Going back to an earlier version takes one click and no rebuild.
- No servers, no build pipeline
- connic builds and runs each release from GitHub, GitLab, Bitbucket or straight from the command line.
- Back to the last good version in one click
- Activate an earlier release and connic switches to it with the connectors and settings it shipped with.
- A clean handover
- A run finishes on the version it started on. Runs still waiting start on the new one.
Limits, retries and concurrency sit in the agent file.
Each setting ships with the release and comes back with a rollback. This file runs the parcel updates shown above.
version: "1.0"
name: parcel-updates
description: Writes delivery updates from scanner events
type: llm
model: connic/gpt-5.6-terra
fallback_model: connic/gpt-5.6-luna
retry_options:
attempts: 3
max_concurrent_runs: 20
timeout: 120
concurrency:
key: data.parcel_id
on_conflict: coalesce
session:
key: context.parcel_id
ttl: 604800fallback_model: connic/gpt-5.6-lunafallback_model, retry_options
When connic/gpt-5.6-terra fails a request, connic/gpt-5.6-luna takes over with up to three attempts. Finished tool calls keep their results.
max_concurrent_runs
At most 20 runs of this agent at once, within your plan's limit.
timeout
A run stops after 120 seconds. The ceiling is your plan's run timeout.
concurrency
One active run per parcel. A newer scan replaces the waiting run, so a burst ends in one run with the latest data.
session
Each parcel keeps its conversation until it has been quiet for seven days.