Advanced

Everstack's agent runtime includes four advanced capabilities that extend what a single agent can do. Each is opt-in through agent configuration, so you only pay for the complexity you need.

Job delegation

Job delegation lets an agent hand off work to a sub-agent asynchronously. The parent agent continues its own work and checks back on the result later.

This is useful when:

a task can be broken into independent pieces that run in parallel
the parent needs to stay responsive while a long-running sub-task completes
you want to specialize agents for different domains and compose them

The parent uses delegate_job to queue structured work for another agent, then calls check_job to poll for completion. Results flow back through the JobResultCh channel into the parent's next loop iteration.

Configure with spawn.async: true and spawn.maxConcurrentJobs in the agent config.

Forking

Forking creates independent branches of an agent's execution. Each fork starts with a copy of the current conversation context and runs autonomously from that point forward.

Use forks when:

you want to explore multiple approaches to a problem simultaneously
a decision point has several viable paths and you want to evaluate all of them
you need to test different strategies before committing to one

Fork results are collected and returned to the originating agent for synthesis.

Configure with fork.enabled: true in the agent config.

Context compaction

Long agent sessions accumulate conversation history that eventually exceeds the model's context window. Context compaction manages this automatically using a three-tier system.

The tiers activate based on how full the context window is:

Background (80% utilization): summarizes the oldest 30% of non-system messages using a lightweight summarization model
Aggressive (85% utilization): summarizes the oldest 60% of non-system messages
Emergency (95% utilization): hard-truncates to the system prompt plus the last 20 user messages, without an LLM call

The summarization model is configurable (defaults to a fast, inexpensive model). System messages are always preserved across all tiers.

Compaction is transparent to the agent. It sees the summarized context as part of its normal conversation history, so it can reference earlier work without knowing a summary replaced the raw messages.

Configure with monitor.enabled: true in the agent config.

Digest bulletins

When multiple agents run concurrently in the same workspace, digest bulletins keep each agent informed about what the others are doing. A DigestManager aggregates activity across all active sessions and injects a short bulletin into each agent's context at the start of its next loop iteration.

This is valuable when:

several agents work on related parts of a larger task
coordination matters but direct message passing would create too much overhead
you want situational awareness without tight coupling between agents

Bulletins are lightweight summaries, not full message histories. They give each agent enough context to avoid duplicating work or conflicting with another agent's actions.

Configure with digest.enabled: true in the agent config.

Configuration summary

All four features are opt-in. Enable them individually based on your use case:

spawn.async: true and spawn.maxConcurrentJobs: N for job delegation
fork.enabled: true for forking
monitor.enabled: true for context compaction
digest.enabled: true for digest bulletins

These features compose well together. A persistent agent might use all four: delegating jobs to specialized sub-agents, forking to explore alternatives, compacting its own context as sessions grow long, and receiving digests about what its peers are doing.

Next steps

Review Tools for the delegation and fork tool interfaces.
See Sessions for how these features affect session state and streaming events.
Use the Agents API to configure these features programmatically.