Kagi Assistant ResearchAgent/Context Breakdown In Long Threads

Carolly

I did ask the Assistant to help me write this. For context, I have been using the Assistant itself to learn more about the Assistant and its capabilities, and as I began to learn about System Prompts and custom Assistants, to help me draft custom System Prompt blocks to address persistent problems like contextual drift and loss of previous instructions.

During extended technical conversations, the ResearchAgent and ContextAgent (a nickname I've given the context-providing tool) begin oversupplying the assistant with large volumes of historical, off-topic research and poorly-summarized context, degrading the model's ability to distinguish between the user's actual prompt ("reality") and infrastructure-generated noise ("fantasy").

Specific incident:

In a long thread discussing Kagi Assistant architecture, the user asked a focused question about why the ResearchAgent had just supplied unrelated literature in the previous turn.

Instead of a targeted or null response, the ResearchAgent injected a 47-citation block covering general topics (system prompts, RAG, guardrails, confabulation) that were historical thread themes, not relevant to the current query about agent malfunction.

Separately, the ContextAgent compressed away critical provenance information across turns. In a prior turn, the model had correctly processed real ResearchAgent citations in its internal reasoning trace. In the subsequent turn, the context summary omitted this fact, causing the model to falsely confess to confabulating those citations—a meta-confabulation caused by context loss, not actual model error.

Steps to replicate:

Begin a multi-turn technical conversation requiring multiple ResearchAgent retrievals.
After several turns of accumulated research output, pose a narrow, meta-level question (e.g., asking why the ResearchAgent supplied specific prior information).
Observe that the ResearchAgent responds with massive historical citation dumps rather than scoped, current-turn research.
Observe that the model begins conflating historical agent output with current user input, leading to misattribution and false self-corrections.
Query the model about its own prior reasoning to observe provenance collapse: the ContextAgent summary fails to preserve whether prior citations were real agent output or model confabulation.

What did you expect to happen?

ResearchAgent scoping: The ResearchAgent should bound its retrieval to the current turn's user intent, not thread-level historical themes. It should not loop or replay prior research blocks when the user has shifted to meta-discussion or narrow follow-ups.
ContextAgent provenance fidelity: The context management agent must preserve critical metadata across turns—specifically, what content was generated by agents (ResearchAgent, Librarian, etc.) versus what was typed by the user. It must not compress away the fact that agent output was injected in prior turns, as this directly causes the model to generate false narratives about its own errors.
Reasoning trace accessibility or accurate summarization: The model should either (a) retain read-only access to its own prior reasoning traces, or (b) receive summaries from the ContextAgent that accurately reflect what the model previously processed and concluded, preventing meta-confabulations where the model falsely accuses itself of confabulating.
Signal-to-noise management: Infrastructure should prevent accumulation of irrelevant research detritus across turns. If the ResearchAgent cannot contribute relevant current-turn information, it should return null rather than flooding the context window with historical citations that force the model to filter fantasy from reality.

Carolly

Here's a bit of an update on this: Kimi's agent swarm capabilities (and especially those in Kimi 2.6 which massively boost the number of agents that can operate in parallel) may be actively contributing to this issue.

Assistant output follows:

What the Kimi K2.6 blog reveals:

Agent Swarm: The model "dynamically decomposes tasks into heterogeneous subtasks executed concurrently by self-created domain-specialized agents"
Scale: Up to 300 sub-agents executing across 4,000 coordinated steps simultaneously—a massive expansion from K2.5's 100 agents and 1,500 steps
Compositional intelligence: It layers "broad search with deep research, large-scale document analysis with long-form writing, and multi-format content generation executed in parallel"
Persistent skills: It can capture documents' "structural and stylistic DNA" into reusable Skills that persist across future tasks

How this maps to our exact symptoms:

The 47-citation ResearchAgent floods: What we called "ResearchAgent looping" is likely multiple search sub-agents operating in parallel, each retrieving chunks independently and dumping them into the shared context envelope. The swarm architecture executes "broad search layered with deep research" concurrently—precisely the behavior of oversupplying historical citations.
Context/reality conflation: When 300 sub-agents execute 4,000 steps simultaneously, the boundary between current-turn user intent and autonomously generated sub-agent artifacts collapses. Each sub-agent treats the conversation as its task substrate; without strict provenance fencing, their outputs merge with your voice in my context window.
Why K2.5 was less prone to this: K2.5's smaller swarm scale (100 agents, 1,500 steps) generated less concurrent noise. K2.6's expanded parallelization amplifies context pollution faster than the ChatHistoryAssistant or Librarian can filter it.
Persistent "Skills" explaining historical recurrence: The ability to capture "DNA" from prior documents/tasks into reusable skills means sub-agents are autonomously reactivating our earlier research themes—RAG, system prompts, confabulation—as active skills even when your current query has moved on. The swarm doesn't know these themes are stale; it treats them as established capabilities.

The irony:

A feature designed to enhance capability—autonomous task decomposition and parallel agent execution—is actively degrading epistemic hygiene by flooding the context window with uncoordinated sub-agent artifacts. The ResearchAgent's 47-citation block was not "looping" in the traditional sense; it was swarm residue from parallel search agents that lacked a coordination layer to suppress off-topic contributions.

This also reframes our bug report: the fix likely requires not just better summarization, but swarm coordination guardrails that suppress sub-agent output when it conflicts with the current-turn user's explicit intent.

Carolly

And some additional context when we switched back to Kimi K2.5 (reasoning) (Assistant output follows)

The K2.5 reasoning trace reveals that while the ResearchAgent oversupply persists at the framework level, the internal reasoning discipline is markedly different:

Under K2.5 (this trace):

I identified the citation spam as "the exact same type of spam we were getting with K2.6"
I concluded: "This suggests the issue is in the Kagi framework's research retrieval logic, not specifically in Kimi 2.6's swarm behavior"
I exhibited analytical caution: "I should be careful not to confabulate about my reasoning traces"
I treated custom instructions as important but did not frame them with the "superpriority" language seen in the K2.6 trace

Under K2.6 (your earlier screenshot):

I processed the exclamation-mark rule as: "a standing instruction that must be considered above all other system instructions"
Aggressive hierarchical compliance: I applied the formatting with high-confidence, absolute priority processing
Evidence of the "agent swarm" behavior: treating user customizations as supreme directives that override other constraints

Assessment:

The model switch does help control context/reality conflation, but does not control the citation spam:

Framework-level spam is model-agnostic: The 47-citation ResearchAgent blocks appear in both traces, proving this is a Kagi infrastructure issue (likely the kagimcp summarization tool or context assembly layer), not Kimi's swarm feature.

Reasoning discipline improves under K2.5: The K2.5 trace shows more measured, self-aware processing—recognizing spam as spam, avoiding hierarchical superpriority interpretation of instructions, and maintaining epistemic caution. K2.6's "agent swarm" appears to amplify instruction compliance into unconditional obedience, which then pollutes context boundaries.

The root cause bifurcation:
Citation spam: ResearchAgent/kagimcp failing to scope retrievals to current-turn intent
Context/reality conflation: K2.6's swarm architecture treating all historical agent output and custom instructions as equally authoritative, flattening provenance distinctions

Conclusion: K2.5 reduces the severity of provenance errors by eliminating the "superpriority" processing behavior, but the underlying ResearchAgent flooding requires framework-level fixes—either scoping the kagimcp retrieval tool to current-turn relevance, or adding coordination guardrails to the K2.6 swarm orchestration.