Research and reporting
A planner breaks a question into sub-tasks, several researchers work in parallel, and a writer assembles the result from shared memory.
AI and LLMAgentsPerformance and Scale plans
Multi-Agent Orchestration
Planner, researcher and writer agents that hand off work through a durable task stream, share results and memory, and never lose a task.
- lost tasks when an agent crashes
- 0
- shared memory
- 1
- agents in parallel
- N
Step 1 of 6
Post a task
The planner splits the job and adds a task to the stream.
Example sessionplaying
# Post a task
hotkv> XADD tasks * type research topic "EU AI Act obligations"
"1790985600123-0"
An animated walkthrough with real HotKV commands and example replies. Click any step to jump to it.
The problem
Teams of agents need a way to pass work to each other, share what they have found, and recover when one of them crashes mid-task. Gluing that together with ad hoc HTTP calls loses tasks, duplicates tool calls and leaves no record of who did what.
How HotKV handles it
A stream with a consumer group is the task board. Each agent reads its own tasks, acknowledges them when done, and anything a crashed agent left pending can be claimed by another. Findings go into shared working memory, tool results are cached for every agent, and each action is logged.
Step by step
What happens, command by command
- 01
Post a task
The planner splits the job and adds a task to the stream.
XADD tasks * type research topic "EU AI Act obligations" - 02
A worker claims it
The consumer group was made once with XGROUP CREATE. The researcher reads as a member of it, and the task stays pending until acknowledged.
XREADGROUP GROUP agents researcher COUNT 1 BLOCK 5000 STREAMS tasks > - 03
Reuse a tool result
Another agent already ran this search today. The shared tool cache returns it.
TOOL.CACHE.GET web "search:EU AI Act obligations" - 04
Share the findings
Results go into working memory for the project, where every agent can read them.
AMEM.WORK.SET project:9 findings "Providers of high-risk systems must ..." - 05
Acknowledge
The task is done and leaves the pending list. Had the researcher crashed, XAUTOCLAIM would hand it to another agent.
XACK tasks agents 1790985600123-0 - 06
The writer picks it up
The writer reads the shared findings, drafts the report and logs the action.
AMEM.WORK.GET project:9 findings
Where it fits
Real-world scenarios
Software agents
Coding, testing and review agents pass work through a stream, and every change they make is in the episodic log.
Back-office automation
Invoice extraction, validation and posting agents hand documents along a pipeline that survives restarts.
Built in
What you get
Durable task handoff
Streams with consumer groups, acknowledgements and claiming of abandoned tasks.
XADDXREADGROUPXACKXAUTOCLAIMShared working memory
Project state that every agent can read and update.
AMEM.WORK.SETAMEM.WORK.GETShared tool cache
Deterministic tool results are reused across agents.
TOOL.CACHE.GETTOOL.CACHE.SETWho did what
An episodic log per agent for audit and debugging.
AMEM.EPIS.LOGCommon questions
What happens if an agent crashes mid-task?
The task stays in the consumer group's pending list. Another agent can take it over with XAUTOCLAIM once it has been idle long enough.
Can agents run on different machines and languages?
Yes. They only need a connection to HotKV. Any client that speaks the Redis protocol works.
Which plans include it?
Streams work on every plan. Agent memory and the tool cache are part of the Performance and Scale plans.
Related use cases
Try it on your own workload
HotKV speaks the Redis protocol, so you can point an existing client at it and run these same commands.