Aegis LangChain DeepAgents Demo

This demo shows how Aegis stabilizes runtime behavior in an existing LangChain/DeepAgents workflow without changing the underlying model or redesigning the app.

• Same task and prompts
• Same workflow shape
• Aegis inserted as a runtime control layer

---

What this demo proves

Aegis is used as a scope-first runtime SDK:

• client.auto().llm(...) for model-call stabilization
• client.auto().step(...) for loop/supervisor step stabilization (where applicable)

This repository does not use the legacy plan-first mental model.

---

Aegis SDK surface used

from aegis import AegisClient, AegisConfig
import os

client = AegisClient(
    api_key=os.environ["AEGIS_API_KEY"],
    base_url=os.getenv("AEGIS_BASE_URL"),
    config=AegisConfig(mode="balanced"),
)

result = client.auto().llm(...)

Returned AegisResult values are inspected in the demo outputs, including:

• actions
• trace
• metrics
• used_fallback
• explanation
• scope
• scope_data

---

Environment

Required

• OPENAI_API_KEY
• AEGIS_API_KEY

Optional

• AEGIS_BASE_URL (set explicitly if not using default endpoint)
• DEMO_MODEL (default: openai:gpt-4o-mini)

Example .env

OPENAI_API_KEY=...
AEGIS_API_KEY=...
AEGIS_BASE_URL=http://localhost:8080
DEMO_MODEL=openai:gpt-4o-mini

---

Run

python -m benchmark_v3.run_baseline
python -m benchmark_v3.run_aegis
python -m runners.compare_benchmark_v3

Other runnable variants

python -m runners.run_aegis
python -m runners.run_working_aegis
python -m runners.run_v2_aegis
python -m runners.run_aegis_supervisor

---

What to inspect in outputs

Aegis runs emit:

• aegis_result*.json files (replacing old plan artifacts)
• optional aegis_debug_summary.txt

Inspect:

• actions → applied runtime controls
• used_fallback → whether fallback behavior was triggered
• trace / metrics → decision and execution context

---

Key takeaway

This demo shows how Aegis sits above an existing system to:

• reduce waste
• stabilize behavior
• improve consistency

…without changing the underlying model integration. Without rewriting your system.

About the Benchmark

This demo is designed to reflect a common real-world pattern:

Many AI workflows are built as multi-pass systems
(planner → solver → validator → refine)

Even when the first answer is already correct, these systems often:

• re-run validation
• perform unnecessary refinement
• make extra model calls by default

In this benchmark:

• the baseline represents that standard multi-pass behavior
• Aegis supervises execution at runtime and stops early when safe

Important:

• Aegis does not improve the model’s intelligence
• it does not change the task or inject answers
• it only reduces unnecessary execution

The goal is to measure efficiency, not raw capability.