What's the purpose of these agents? Or how to use them.

[JonasDoe]

I asked Claude:

How useful deem you agents as such https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/09-meta-orchestration/knowledge-synthesizer.md? For me, it looks just like a huge list of buzzwords, I don't really understand how this is helping a dedicated agent.

The response:

Your instinct is correct. This is a system prompt, not an agent — and it's a bad one.

What it actually is: frontmatter (name/description/tools/model) + a giant wall of bullet-point noun phrases fed as a persona prompt to Sonnet. There's no code, no pipeline, no actual pattern-mining logic, no knowledge graph implementation — just prose telling an LLM to pretend it has these capabilities.

Why it's mostly buzzword theater:

1. Fabricated metrics. "Pattern accuracy > 85%", "Knowledge retrieval < 500ms", "50k+ entities" — these numbers appear nowhere as actual measurements. There's no instrumentation, no eval harness, no data source computing them. The model will just hallucinate plausible-sounding numbers in its "Delivery notification" because the prompt tells it to output that format. It's asking the LLM to lie convincingly.

2. No mechanism, only vocabulary. Sections like "Learning mechanisms: Supervised learning / Unsupervised discovery / Reinforcement > learning / Federated learning..." — these are just labels. There's no code implementing any ML training loop; a Claude subagent with tools: Read, Write, Edit, Glob, Grep cannot run federated learning. It's listing techniques it has zero capacity to execute, presumably so the model's output sounds like it considered them.

3. Circular "Communication Protocol." The JSON blobs ("request knowledge context from context manager") assume an orchestration layer that doesn't exist in this repo — no actual message bus, no other agents to call. It's cargo-culting multi-agent-system jargon onto what is, functionally, a single Claude Code subagent invocation with file read/write tools.

4. Tools don't match claims. It claims to "build a knowledge graph," do "predictive insights," "anomaly detection," "causation inference" — with only Read, Write, Edit, Glob, Grep. Those tools can grep log files and write markdown notes. That's it. Everything past "read some files and summarize patterns you notice" is aspirational prose the model can't actually back up.

What's actually salvageable: strip it down to "when invoked, grep through logs/session-history files, look for recurring patterns in errors or successful workflows, write findings to a knowledge.md file." That's maybe 10 lines. The other ~270 lines are keyword-stuffing — likely written (or generated) to look comprehensive/impressive for a public "awesome-list" repo rather than to function well. This is a common failure mode in these subagent marketplace repos: the prompt is optimized for looking sophisticated in a README, not for constraining model behavior toward a concrete, checkable task.

If you want a synthesizer subagent that's actually useful for your Claude Code / OpenCode workflow, the design should be: narrow input (specific log/session file glob), narrow output schema (e.g. 5-10 line JSON of {pattern, evidence, frequency}), and grounding — no "insight relevance > 90%" theater, since there's nothing computing that number.

Am I missing something here? Esp. point 3 makes me wonder whether I should set up some kind of orchestration layer this queries should be fired against.

[necatiozmen]

Thanks for raising this — the critique is fair, and you're not missing anything on point 3: these are plain Claude Code subagents, not nodes in a running orchestration system. There is no message bus to set up. The Communication Protocol JSON blobs were decorative and misleading, implying a runtime that doesn't exist in this repo. Subagents here coordinate only two ways: through shared files and through the orchestrator (agent or human) that invokes them and passes context along. So no, you don't need to build a layer for those queries to fire against.

The rest of the critique lands too. With tools: Read, Write, Edit, Glob, Grep, an agent can search text and write Markdown — it can't run federated learning, build a live knowledge graph, or compute "pattern accuracy > 85%". Those numbers, plus the Identified 342 patterns... improving performance by 23% style notifications, were exactly what you called them: the prompt telling the model to output confident fabricated metrics.

We've reworked the entire 09-meta-orchestration category in #308:

• Removed all fabricated metrics (numeric-target checklists, progress JSON with invented counts, delivery-brag notifications).
• Cut the technique/buzzword blocks the tools can't execute.
• Replaced the circular Communication Protocol with honest file-and-orchestrator coordination, plus grounding rules (cite path:line evidence, don't invent numbers, flag uncertainty).

The 8 files dropped from ~286 lines each to ~66–101, keeping the genuinely useful workflow content. Appreciate the detailed writeup — it made the fix straightforward.