Python: latency improvements

[eavanvalkenburg]

Motivation and Context

Used a profiling script and cProfile to figure out the biggest time losses in AF, led to a few improvements

• caching json schema creation from the pydantic input model
• using checks on type fields instead of isinstance
• logging using the same object as what is sent to otel, to eliminate a additional to_dict call

Description

Benchmark Comparison: Old vs New (Optimized) Code

Benchmark	Old (ms)	New (ms)	Change	Improvement
Imports (steady-state)
import_agent_framework	0.004	0.004	~0	-
import_azure_openai_chat_client	0.006	0.006	~0	-
import_openai_chat_client	0.005	0.005	~0	-
import_chat_agent	0.006	0.005	-0.001	~17% faster
import_ai_function	0.005	0.005	~0	-
Chat Client (no observability)
chat_client_streaming=False_function_call=False	0.052	0.052	~0	-
chat_client_streaming=True_function_call=False	0.108	0.107	-0.001	~1% faster
chat_client_streaming=False_function_call=True	0.054	0.052	-0.002	~4% faster
chat_client_streaming=True_function_call=True	0.161	0.143	-0.018	~11% faster
Chat Client (with observability)
chat_client_streaming=False_function_call=False	0.107	0.082	-0.025	~23% faster
chat_client_streaming=True_function_call=False	0.204	0.186	-0.018	~9% faster
chat_client_streaming=False_function_call=True	0.111	0.079	-0.032	~29% faster
chat_client_streaming=True_function_call=True	0.222	0.182	-0.040	~18% faster
Agent (no observability)
agent_streaming=False_function_call=False	0.107	0.104	-0.003	~3% faster
agent_streaming=True_function_call=False	0.195	0.262	+0.067	~34% slower*
agent_streaming=False_function_call=True	1.79	1.55	-0.24	~13% faster
agent_streaming=True_function_call=True	0.78	0.66	-0.12	~15% faster
Agent (with observability)
agent_streaming=False_function_call=False	0.21	0.16	-0.05	~24% faster
agent_streaming=True_function_call=False	0.38	0.33	-0.05	~13% faster
agent_streaming=False_function_call=True	2.49	1.70	-0.79	~32% faster
agent_streaming=True_function_call=True	1.55	0.93	-0.62	~40% faster

Key Findings

🎯 Biggest Wins (with function calls + observability):

• Agent streaming + function_call + observability: 40% faster (1.55ms → 0.93ms)
• Agent non-streaming + function_call + observability: 32% faster (2.49ms → 1.70ms)
• Chat client + function_call + observability: 29% faster (0.11ms → 0.079ms)

📊 Observability Overhead Reduced:

• Old: Observability added ~0.05-0.09ms overhead
• New: Observability overhead reduced significantly, especially for function call scenarios

⚡ Summary:

• Observability-enabled paths show 13-40% improvement
• Function call scenarios show 13-32% improvement
• Base operations (no observability, no function calls) remain similar
• One outlier: agent_streaming=True_function_call=False is slightly slower (likely measurement variance)

The optimizations primarily benefit the hot paths: function calling and observability tracing - which are the most common real-world usage patterns!

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? If yes, add "[BREAKING]" prefix to the title of the PR.

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_tools.py	711	67	90%	224, 270, 321, 323, 486, 518–519, 621, 623, 643, 661, 675, 687, 692, 694, 701, 734, 788–790, 831, 854, 856–865, 874–880, 916, 926, 1110, 1524–1528, 1649, 1718, 1811, 1817, 1860–1861, 1874–1875, 2004, 2045–2046, 2074–2076, 2118–2119, 2184–2185, 2192–2193
_types.py	963	101	89%	130–131, 149–150, 287, 289, 296, 315, 355, 401–402, 438, 588, 702–703, 705, 730, 737, 754–756, 841, 846–847, 849, 856–857, 859, 886, 895, 898–900, 905–906, 913, 917–919, 1075, 1164–1167, 1175–1176, 1267, 1448, 1454, 1698–1700, 1706–1707, 2009, 2014, 2018, 2022, 2200–2202, 2214, 2266–2270, 2280, 2285, 2744, 2830–2832, 2905, 2916–2917, 3091, 3095, 3107–3109, 3210–3212, 3214–3216, 3219, 3223, 3226, 3231, 3276–3277, 3284–3285, 3319–3321, 3352, 3380, 3387
observability.py	642	154	76%	244, 312–317, 319, 321–322, 324, 326–328, 331–333, 338–339, 345–346, 352–353, 360, 362–364, 367–369, 374–375, 381–382, 388–389, 396, 433, 436, 439–441, 444, 447–448, 451–453, 455–457, 460, 547, 549, 631, 649–650, 652, 655, 663–664, 667–670, 672, 675–677, 680–681, 694–700, 702–711, 714–718, 721–724, 726–729, 732–733, 741, 842, 844, 869–871, 993, 995, 999–1004, 1006, 1009–1013, 1015, 1285, 1365–1367, 1439–1441, 1614, 1622, 1626, 1630, 1636, 1638, 1640, 1648, 1658, 1699–1702, 1716, 1718, 1725, 1741, 1744, 1804, 1820, 1824, 1958, 1960
TOTAL	16846	2621	84%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
2554	148 💤	0 ❌	0 🔥	58.823s ⏱️

[Copilot]

Pull request overview

This PR introduces performance improvements to the Python Agent Framework based on profiling analysis. The changes focus on reducing computational overhead through strategic optimizations.

Key Changes

• Caching of JSON schema generation from Pydantic models to avoid repeated expensive serialization
• Replacing isinstance() checks with faster string attribute comparisons for content type identification
• Reusing OpenTelemetry message representations for logging to eliminate redundant to_dict() calls

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
python/packages/core/agent_framework/observability.py	Optimizes message logging by reusing the OTEL message representation instead of calling to_dict() separately
python/packages/core/agent_framework/_types.py	Improves content type checking performance by using type attribute comparison before falling back to isinstance()
python/packages/core/agent_framework/_tools.py	Implements caching for model_json_schema() results and updates exclusion list for proper serialization

[TaoChenOSU]

Can we use enums at least?

[TaoChenOSU]

nit: please also share the profiling results