- Every exit is recorded with its exit code and time
- Restarts back off so a crash loop cannot thrash the server
- Start, stop and restart from the console or the API
Platform
Everything an agent needs
to run unattended.
Byte Roast gives every agent its own server and a managed stack on top. Here is what ships at launch, and the agent-specific features we are building next.
{{countText}}
- Never written into your code or your image
- Rotate a key without redeploying
- Each agent only sees the secrets you assign to it
- Filter by level to find the failing tool call
- Logs survive restarts
- Download a time range for a bug report
- Private network between your servers in a region
- Inbound and outbound firewall rules per agent
- Load balancers with health checks and reserved IPs
- Restore a server to an earlier state in a few clicks
- Snapshots are kept until you delete them
- Backups cost 20% of the server price
- Alerts by email, Slack or webhook
- Alert when an agent restarts more than you expect
- One activity log for every change in the workspace
- Support, research, Slack, coding, webhook, eval and scheduled agents
- Bring any Dockerfile
- Change the code, keep the supervision, secrets and logs
- Owner, admin, member and billing roles
- Read or read-write API keys that expire
- Two-factor sign-in and session management
Agents in a loop can burn through model credits overnight. This caps the damage.
- Outbound calls to model providers pass through a metering proxy
- Warn at 80% of budget, pause the agent at 100%
- See spend per agent next to its server cost
An agent that can browse can be talked into calling the wrong place.
- Allowlist per agent, edited without a redeploy
- Every outbound request recorded with destination and status
- Blocks prompt-injected attempts to send data elsewhere
Logs tell you that it failed. A trace shows you why.
- OpenTelemetry-compatible, so your own tools can read it
- Open the exact step where a run went wrong
- Replay a failed run against a fixed version
The worst agent failures do not crash. They just stop making progress.
- Restart when the heartbeat stops, not only when the process exits
- Flag runs that repeat the same tool call
- Alert with the last log lines attached
Let agents do real work without giving them the last word.
- Mark tools as needing approval, such as refunds or deletes
- Approve or reject from a message, with a timeout
- Every decision kept in the audit log
- Cron schedules with time zones
- Signed webhook endpoints per agent
- A managed queue with retries and a dead-letter list
Some agents only work a few hours a day. Hourly billing should reflect that.
- Sleep after a quiet period you choose
- Wake on a schedule, a webhook or a click
- Snapshot storage billed per GB while asleep
- Fresh environment per run, deleted afterwards
- No access to your secrets or private network
- Time and memory limits per run
- Connect a repository and pick a branch
- Keep the last ten releases
- Preview agents for pull requests
- Vector search for embeddings and long-term memory
- Reachable only from your agents
- Daily backups and point-in-time restore
Nothing matches these filters.
Shape the roadmap
Which one should we build first?
Tell us what your agents need. Early-access customers get a say in the order.