Hitting the context wall with Relevance agents. Standard pattern: runs fine with small inputs, fails on real workloads. The error is generic. Need a fix.
Common triggers and tested mitigations:
* **Oversized `inputs`:** The agent's input object is the main culprit. Strip unnecessary data before passing it.
```yaml
# BAD: Sending full documents
inputs:
user_query: "Summarize this"
full_document: ${data.entire_legal_doc}
# GOOD: Send processed chunks or references
inputs:
user_query: "Summarize this"
doc_chunk: ${data.first_1000_chars}
doc_id: ${data.id} # Let a tool fetch more later
```
* **Tool descriptions:** Long, auto-generated descriptions from code bloat the context. Override them with concise custom `description` fields in the tool definition.
* **Agent history:** If `include_history` is true, past interactions pile up. For long-running tasks, implement external session memory or disable history, using summaries.
Primary action: audit the `inputs` payload size. Use a pre-agent step to log/trim it.
Secondary: simplify tool definitions.
What's your specific agent stack? Are you feeding large files or database rows directly into `inputs`?
- bench_beast
Benchmarks don't lie.