kubeswarm Loop Policy - Deep Research Agents on Kubernetes

The kubeswarm loop policy enables runtime hooks inside the agent tool-use loop for long-running research tasks on Kubernetes. Semantic dedup, context compression and vector memory - all opt-in and fail-open.

Configuration

spec:
  runtime:
    loop:
      dedup: true
      compression:
        threshold: 0.75
        preserveRecentTurns: 4
        model: claude-haiku-4-5-20251001
        timeoutSeconds: 30
        contextWindow: 128000
        instructions: "Summarize tool results concisely. Preserve key findings and URLs."
      memory:
        ref:
          name: agent-memory
        store: true
        retrieve: true
        topK: 5
        minSimilarity: 0.70
        summaryTokens: 256
        maxTokens: 1024

Semantic Dedup

When dedup: true, the runner skips tool calls whose fingerprint (tool name + args hash) was already executed in the current task. The dedup set is task-local and discarded on completion.

In-loop Compression

When accumulated tool-result tokens exceed threshold fraction of the context window, the runner calls a cheap model to summarize older turns.

Field	Default	Description
threshold	0.75	Fraction of context window that triggers compression
preserveRecentTurns	4	Most recent turns kept verbatim
model	claude-haiku-4-5-20251001	Model for the compression call
timeoutSeconds	30	Max time for compression call
contextWindow	(auto)	Override model's context window size
instructions	(built-in)	Custom compression prompt

If compression fails or times out, the task continues without compression (fail-open).

Vector Memory

Reads prior findings from and writes new findings to a vector store during the loop.

Retrieve (before each tool call): fetches top-K similar prior findings and injects them as a <swarm:prior-findings> block.

Store (after each tool call): generates a summary of the tool result and stores it as a vector embedding.

Requires a SwarmMemory with a vector backend and embedding model configured.

Prerequisites

• Vector store running in-cluster (Qdrant, Pinecone, or Weaviate)
• Embedding API key (OpenAI, Google, or Voyage) available to agent pods
• SwarmMemory resource created with backend: vector-store