Context Engineering System
OmniCoreAgent implements state-of-the-art context engineering inspired by patterns from Anthropic and Cursor. This dual-layer approach ensures your agents never hit token limits — even during marathon coding sessions or multi-step research tasks.
How the Two Layers Work Together
| Layer | Scope | What It Manages | When It Triggers |
|---|
| Context Management | Agent loop messages | User/Assistant conversation, tool call history | When context exceeds threshold_percent of limit |
| Tool Offloading | Individual tool responses | Large API responses, file contents, search results | When response exceeds threshold_tokens |
Layer 1: Agent Loop Context Management
Prevent token exhaustion during long-running tasks with automatic context management. When enabled, the agent monitors context size and applies truncation or summarization when thresholds are exceeded.
agent_config = {
"context_management": {
"enabled": True,
"mode": "token_budget", # or "sliding_window"
"value": 100000, # Max tokens (token_budget) or max messages (sliding_window)
"threshold_percent": 75, # Trigger at 75% of limit
"strategy": "summarize_and_truncate", # or "truncate"
"preserve_recent": 4, # Always keep last N messages
}
}
Modes
| Mode | Description | Best For |
|---|
token_budget | Manage by total token count | Cost control, API limits |
sliding_window | Manage by message count | Predictable context size |
Strategies
| Strategy | Behavior | Trade-off |
|---|
truncate | Drop oldest messages | Fast, no extra LLM calls |
summarize_and_truncate | Summarize then drop | Preserves context, adds latency |
Large tool responses are automatically saved to files, with only a preview in context. The agent can retrieve full content on demand using built-in tools.
agent_config = {
"tool_offload": {
"enabled": True,
"threshold_tokens": 500, # Offload responses > 500 tokens
"max_preview_tokens": 150, # Show first 150 tokens in context
"storage_dir": "workspace/artifacts"
}
}
Token Savings
| Tool Response | Without Offloading | With Offloading | Savings |
|---|
| Web search (50 results) | ~10,000 tokens | ~200 tokens | 98% |
| Large API response | ~5,000 tokens | ~150 tokens | 97% |
| File read (1000 lines) | ~8,000 tokens | ~200 tokens | 97% |
| Tool | Purpose |
|---|
read_artifact(artifact_id) | Read full content when needed |
tail_artifact(artifact_id, lines) | Read last N lines (great for logs) |
search_artifact(artifact_id, query) | Search within large responses |
list_artifacts() | See all offloaded data in current session |
Combined Power
Enable both for maximum efficiency:
agent = OmniCoreAgent(
name="research_agent",
agent_config={
"context_management": {"enabled": True, "strategy": "summarize_and_truncate"},
"tool_offload": {"enabled": True, "threshold_tokens": 500}
}
)
# Result: Agents that can run indefinitely without token exhaustion
Enable both layers for long-running tasks (research, multi-step workflows) where context or tool responses can grow unbounded.