Context Management for Agents¶
28 Agent Skills and a callable runtime (MCP, REST, SDKs) for context engineering across GitHub Copilot, Cursor, Kiro, Antigravity, and Amazon Q.
Why this project¶
Agent quality depends on what fits in the context window. This collection teaches progressive disclosure, compaction, masking, and budgeting — with measured savings on realistic fixtures and a Skill Router that picks the right guidance for a task.
Try it in 60 seconds¶
git clone https://github.com/navendubrajesh/context-management-for-agents.git
cd context-management-for-agents
pip install -e runtime/core -e runtime/mcp
python examples/demo/run_demo.py
You will see the router select context-compression for a handoff task and print before/after token counts from an offline primitive.
What's included¶
| Layer | Location | Purpose |
|---|---|---|
| Skills | skills/ |
28 SKILL.md files (platform + foundational) |
| Runtime | runtime/ |
Router, loader, MCP, REST, Python/TS SDKs |
| Primitives | runtime/core/context_skills/primitives/ |
Masking, compaction, budgeter, pipeline |
| Validation | researcher/scripts/ |
Four deterministic quality gates |
| Enterprise | control-plane/, deploy/ |
IAM, tenancy, policy, audit (Phase 3) |
Documentation map¶
- Quickstart — install, demo, MCP, REST
- Concepts — context engineering techniques
- Skills catalog — auto-generated index
- Runtime reference — MCP tools and REST endpoints
- Benchmarks — measured savings (with caveats)
- Enterprise landing — overview and links
- Enterprise deployment — regulated deployment patterns
Honest claims¶
Measured savings (~93% on handoff fixture, ~93% aggregate on project fixtures) are re-measured on this repository's test fixtures — not theoretical. See Benchmarks. Cost figures in the runtime are estimated; pricing is operator-configured.