Python: [BREAKING] simplify ag-ui run logic, fix mcp bugs, fix anthropic client issues in ag-ui

[moonbox3]

Motivation and Context

This refactor restructures AG-UI orchestration to reduce bespoke state handling and make the flow maintainable. The existing code was growing in complexity and brittleness; this change consolidates orchestration, simplifies state transitions, and removes ad hoc management paths.

Important

The only breaking change here is we're removing the confirmation strategies - no need for this custom behavior. The confirmation_strategy arg is no longer passed into AgentFrameworkAgent.

Description

• Fixes Python: [Bug]: AG-UI integration with Anthropic Foundry Agent throws error #3278
• Fixes Python: [AG-UI] MCP tools not executing when using human-in-the-loop approval #3297
• Fixes Python: [AG-UI] After running the MCP tools, a warning appears on the screen #3298
• Fixes Python: [Bug]: Missing fields in AGUIRequest cause client_tools loss in DefaultOrchestrator #3350
• Fixes Python: [AG-UI] Error occurs when using with_plan_review in Magentic workflow #3331

Testing:

• All samples from AG-UI dojo were tested with the AzureOpenAIClient, AnthropicClient, and the AzureAIClient
• Multiple MCP tools were tested
• Human-in-the-loop approvals with MCP tools were tested
• Unit test coverage > 85% for current package

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/ag-ui/agent_framework_ag_ui
_agent.py	36	1	97%	62
_client.py	140	11	92%	37, 88–92, 238, 268, 423–425
_endpoint.py	46	1	97%	92
_message_adapters.py	442	89	79%	65, 75–76, 85–88, 117–118, 121, 125–127, 136–138, 147–158, 160–162, 192, 205–206, 216–217, 254, 257, 259, 262, 265, 281, 298, 320, 351, 356, 367–368, 419, 435–436, 496–499, 501, 507, 515–516, 518, 522–525, 538, 627–630, 632, 697, 732–734, 736–739, 742–743, 745, 751, 754, 756, 759, 761, 767–768, 770
_run.py	432	121	71%	151–158, 301, 320–321, 336–337, 348, 351–352, 354–355, 357–359, 369, 379–382, 386–388, 390, 400, 403–406, 408–409, 412–418, 421–423, 426, 442–444, 451, 457–458, 460–461, 475–481, 492, 505, 507–508, 542–543, 595–597, 609–611, 635, 640–642, 758, 769–770, 777, 824–826, 841, 847, 855, 857, 893–899, 902–905, 907–916, 919, 926–927, 932, 938–940, 953–955
_types.py	36	1	97%	12
packages/ag-ui/agent_framework_ag_ui/_orchestration
_helpers.py	107	0	100%
_tooling.py	57	0	100%
packages/anthropic/agent_framework_anthropic
_chat_client.py	362	151	58%	51, 362, 385, 387, 402, 424–427, 436, 438, 469–473, 475, 477–478, 480, 485–486, 488, 521–522, 531, 533–534, 539, 556–557, 599, 614, 618–619, 635, 644, 646, 650–651, 694–696, 698, 711–712, 719–721, 725–727, 731–734, 745, 747, 769, 779, 801–807, 814–815, 823–824, 832–835, 842–843, 849–850, 856–857, 863, 871–873, 877, 884–885, 891–892, 898–899, 905, 913–916, 923–924, 943, 950–951, 970, 992, 994, 1003–1004, 1010, 1032–1033, 1039–1040, 1049–1059, 1066–1072, 1079–1085, 1092–1101, 1108–1111
packages/core/agent_framework
_agents.py	299	52	82%	51, 90, 98, 101, 104, 406–408, 454, 636, 805, 808–810, 928–930, 935–937, 1038, 1079, 1081, 1090–1095, 1101, 1103, 1113–1114, 1121, 1123–1124, 1132–1136, 1144–1145, 1147, 1152, 1154, 1188, 1233–1234, 1236, 1238, 1249
_tools.py	770	74	90%	226, 272, 323, 325, 353, 523, 555–556, 658, 660, 680, 698, 712, 724, 729, 731, 738, 771, 827–829, 870, 895–904, 910–919, 955, 965, 1206, 1543, 1629–1633, 1772, 1774, 1840, 1932, 1938, 1980–1981, 1994–1995, 2038, 2122, 2160–2161, 2200–2202, 2240–2241, 2251, 2308–2309, 2316–2317
TOTAL	17006	2649	84%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3229	213 💤	0 ❌	0 🔥	1m 4s ⏱️

[Copilot]

Pull request overview

This PR performs a major refactoring of the AG-UI orchestration system to simplify the run logic and fix several bugs. The changes consolidate orchestration into a single linear flow in _run.py, removing the complex orchestrator pattern and confirmation strategies.

Changes:

• Simplified orchestration by consolidating into a single run_agent_stream() function in _run.py
• Fixed Anthropic client bug where empty text content was causing API errors
• Fixed MCP tool approval handling by ensuring approval tools are properly passed to the execution handler
• Removed confirmation strategies (BREAKING CHANGE)

Reviewed changes

Copilot reviewed 38 out of 38 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
python/packages/core/agent_framework/_tools.py	Added debug logging and fixed conversation_id filtering in tool invocation
python/packages/core/agent_framework/_agents.py	Added tool deduplication by name when merging options
python/packages/anthropic/agent_framework_anthropic/_chat_client.py	Fixed empty text content handling and added conversation_id to ignored options
python/packages/ag-ui/agent_framework_ag_ui/_run.py	New 949-line orchestration implementation replacing orchestrators
python/packages/ag-ui/agent_framework_ag_ui/_orchestration/_tooling.py	Fixed approval tool detection to ensure tools are passed when needed
python/packages/ag-ui/agent_framework_ag_ui/_agent.py	Simplified to use new run_agent_stream() function
python/packages/ag-ui/agent_framework_ag_ui/__init__.py	Removed confirmation strategy exports
python/packages/ag-ui/agent_framework_ag_ui_examples/agents/	Removed confirmation_strategy parameters from agent examples
Multiple test files	Removed obsolete tests and added new tests for simplified architecture