mirror of
https://github.com/furyhawk/deepresearch.git
synced 2026-07-21 02:06:48 +00:00
Add report on Agentic AI breakthroughs in Q2 2026
This report outlines significant advancements in Agentic AI, including the transition to Full-Horizon Planning, enhanced tool interaction, the concept of Memory-as-Ontology, and developments in multi-agent systems. Key breakthroughs and future outlooks are discussed, providing a comprehensive overview of the evolving landscape in the field.
This commit is contained in:
@@ -32,7 +32,7 @@ from pydantic_deep import (
|
||||
from pydantic_deep.toolsets.plan import create_plan_toolset
|
||||
from pydantic_deep.types import SubAgentConfig
|
||||
|
||||
from .config import SKILLS_DIR, get_model
|
||||
from .config import SKILLS_DIR, USE_RATE_LIMITER, get_model
|
||||
from .middleware import RateLimitRetryCapability
|
||||
from .prompts import RESEARCH_PROMPT
|
||||
from .todo_toolset import ForgiveWriteTodosCapability
|
||||
@@ -423,7 +423,8 @@ def create_research_agent(
|
||||
skills=PROGRAMMATIC_SKILLS,
|
||||
skill_directories=[{"path": str(SKILLS_DIR), "recursive": True}],
|
||||
hooks=HOOKS,
|
||||
middleware=[ForgiveWriteTodosCapability(), RateLimitRetryCapability(), *(middleware or [])],
|
||||
middleware=[ForgiveWriteTodosCapability(), *(middleware or [])]
|
||||
+ ([RateLimitRetryCapability()] if USE_RATE_LIMITER else []),
|
||||
context_manager=True,
|
||||
context_manager_max_tokens=200_000,
|
||||
patch_tool_calls=True,
|
||||
|
||||
@@ -652,7 +652,7 @@ async def lifespan(app: FastAPI):
|
||||
)
|
||||
session_manager.start_cleanup_loop(interval=300)
|
||||
|
||||
def _print_banner(servers: list) -> None:
|
||||
def _print_banner(servers: list, middleware: list | None = None) -> None:
|
||||
names = [getattr(s, "prefix", "unknown") for s in servers]
|
||||
print("=" * 60)
|
||||
print("DeepResearch — Full-Featured Research Agent")
|
||||
@@ -662,14 +662,16 @@ async def lifespan(app: FastAPI):
|
||||
print(f" Workspaces : {WORKSPACES_DIR}")
|
||||
print(" Runtime : python-datascience")
|
||||
print(" Hooks : audit_logger, safety_gate")
|
||||
print(" Capabilities : AuditCapability, PermissionCapability")
|
||||
if middleware:
|
||||
mw_names = [type(m).__name__ for m in middleware]
|
||||
print(f" Middleware : {', '.join(mw_names)}")
|
||||
print(" Subagents : code-reviewer, general-purpose + dynamic factory")
|
||||
print(" Execute : enabled (human-in-the-loop)")
|
||||
print(" Image support : enabled")
|
||||
print(" Plan mode : enabled (ask_user)")
|
||||
print("=" * 60)
|
||||
|
||||
_print_banner(mcp_servers)
|
||||
_print_banner(mcp_servers, middleware=[audit_cap, permission_cap])
|
||||
|
||||
# Start MCP server connections — retry loop that drops failing servers
|
||||
remaining = list(mcp_servers)
|
||||
@@ -688,7 +690,7 @@ async def lifespan(app: FastAPI):
|
||||
)
|
||||
remaining = [s for s in remaining if getattr(s, "prefix", "") not in failed]
|
||||
agent = create_research_agent(mcp_servers=remaining, middleware=[audit_cap, permission_cap])
|
||||
_print_banner(remaining)
|
||||
_print_banner(remaining, middleware=[audit_cap, permission_cap])
|
||||
else:
|
||||
break
|
||||
|
||||
|
||||
@@ -27,6 +27,7 @@ STATIC_DIR = APP_DIR / "static"
|
||||
MODEL_NAME: str = os.getenv("MODEL_NAME", "openai-responses:o4-mini")
|
||||
LLM_BASE_URL: str = os.getenv("LLM_BASE_URL", "http://localhost:8011/v1")
|
||||
LLM_API_KEY: str | None = os.getenv("LLM_API_KEY")
|
||||
USE_RATE_LIMITER: bool = (LLM_API_KEY or "no-key-required") != "no-key-required"
|
||||
|
||||
|
||||
def get_model() -> str | OpenAIModel:
|
||||
@@ -111,12 +112,13 @@ def create_mcp_servers() -> list[AbstractToolset]:
|
||||
|
||||
# Jina AI Reader — converts any URL to readable markdown
|
||||
# Requires JINA_API_KEY (free tier available at https://jina.ai/)
|
||||
# MCP endpoint: https://mcp.jina.ai/v1
|
||||
jina_key = os.getenv("JINA_API_KEY")
|
||||
if jina_key:
|
||||
servers.append(
|
||||
PrefixedToolset(
|
||||
MCPToolset(
|
||||
"https://r.jina.ai/mcp",
|
||||
"https://mcp.jina.ai/v1",
|
||||
headers={"Authorization": f"Bearer {jina_key}"},
|
||||
max_retries=3,
|
||||
),
|
||||
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
@@ -1 +1 @@
|
||||
{"session_id": "fa1132cd-34db-4576-aa0c-9ea7c6e8b923", "created_at": "2026-06-17T06:42:28.817230+00:00", "updated_at": "2026-06-17T06:42:28.817230+00:00", "title": "research on latest agentic AI news for the past 3 months sin", "message_count": 0, "todos": [{"active_form": "Researching Reasoning & Planning breakthroughs", "content": "Research advances in long-horizon planning and multi-step reasoning for agents (March-June 2026)", "status": "pending"}, {"active_form": "Researching Tool Use & Interaction capabilities", "content": "Research breakthroughs in complex tool use, web navigation, and cross-platform API integration", "status": "pending"}, {"active_form": "Researching Memory & State Management", "content": "Research progress in persistent agent memory and long-term context management", "status": "pending"}, {"active_form": "Researching Multi-Agent Systems (MAS) trends", "content": "Research trends in multi-agent collaboration, swarms, and hierarchical agent structures", "status": "pending"}, {"active_form": "Synthesizing final report", "content": "Synthesize all research into a high-level overview report in /workspace/report.md", "status": "pending"}]}
|
||||
{"session_id": "fa1132cd-34db-4576-aa0c-9ea7c6e8b923", "created_at": "2026-06-17T06:42:28.817230+00:00", "updated_at": "2026-06-17T08:41:59.925481+00:00", "title": "research on latest agentic AI news for the past 3 months sin", "message_count": 34, "todos": [{"active_form": "Synthesizing final report", "content": "Synthesize findings into a high-level overview report in /workspace/report.md.", "status": "in_progress"}]}
|
||||
@@ -0,0 +1,59 @@
|
||||
# High-Level Overview: Agentic AI Breakthroughs (Q2 2026)
|
||||
|
||||
## Executive Summary
|
||||
The period between March and June 2026 marks a pivotal transition in Agentic AI from "single-action" execution to **Coordination Engineering** and **Full-Horizon Planning**. The industry is moving away from simple tool calling toward robust, autonomous systems capable of managing hundreds of sequential steps, maintaining persistent digital identities (Memory-as-Ontology), and operating within structured multi-agent swarms. Key breakthroughs include the emergence of "Full-Horizon" planning architectures, visual/semantic agentic browsers, constitutional memory governance, and self-evolving coordination protocols.
|
||||
|
||||
---
|
||||
|
||||
## 1. Reasoning & Planning: From Single-Step to Full-Horizon
|
||||
A major architectural shift has occurred in how agents approach complex tasks. The industry is moving away from "Single-Step Horizon" (SH) planning—where an agent plans one action at a time—toward **Full-Horizon (FH) Planning**.
|
||||
|
||||
### Key Breakthroughs:
|
||||
- **Token Efficiency & Accuracy:** FH planning allows agents to generate a complete multi-step plan upfront before execution. For well-defined tasks, this is 2x–4x more token-efficient than SH while maintaining equivalent accuracy.
|
||||
- **Dual-System Architectures:** The **Goal2Skill Framework (April 2026)** popularized the use of high-level VLM planners for "reflective reasoning" and global goal decomposition, paired with low-level executors for specific actions.
|
||||
- **Frontier Model Capabilities:**
|
||||
- **GPT-5.2-high** has demonstrated high reliability in adapting to failures during long-horizon task execution.
|
||||
- **Kimi K2 Thinking** now supports up to 300 sequential tool calls and 4,000 coordinated steps in a single "Preserve-thinking" mode.
|
||||
|
||||
---
|
||||
|
||||
## 2. Tool Interaction & Autonomous Navigation
|
||||
Agents are no longer just "calling functions"; they are interacting with complex software environments as autonomous entities.
|
||||
|
||||
### Key Breakthroughs:
|
||||
- **Agentic Browsers:** Tools like **Perplexity's Comet** and **OpenAI Operator** treat the browser as a navigable environment, using visual and semantic interpretation to "see" and interact with UIs rather than relying solely on brittle DOM-based selectors.
|
||||
- **Coordinated Tool Chains:** New systems can manage hundreds of tool calls (up to 1,500) in a single session, enabling massive file operations and complex database interactions without losing state.
|
||||
- **Standardized Orientation:** The emergence of `llms.txt` at website roots is beginning to standardize how agents discover and navigate site structures during multi-step workflows.
|
||||
|
||||
---
|
||||
|
||||
## 3. Memory & Persistent State: Memory-as-Ontology
|
||||
The most significant conceptual shift in Q2 2026 is the transition from **Memory-as-Tool** (RAG) to **Memory-as-Ontology**. Memory is now viewed as the foundational identity of a persistent digital being.
|
||||
|
||||
### Key Breakthroughs:
|
||||
- **Constitutional Memory Architecture (CMA):** This new framework proposes four governance layers: *Constitution* (safety), *Contract* (system rules), *Adaptation* (personalization), and *Implementation* (technical storage).
|
||||
- **Stability-based Storage:** Instead of just "recency," memories are now organized by stability tiers (High, Mid, Low) to protect identity-critical info while allowing for transient task logs.
|
||||
- **Procedural Memory:** The distinction between episodic ("what happened"), semantic ("what is known"), and procedural memory ("how things are done") has become a standard for building coherent long-term agent behavior.
|
||||
|
||||
---
|
||||
|
||||
## 4. Multi-Agent Systems (MAS) & Swarm Intelligence
|
||||
The focus has shifted from optimizing individual agents to **Coordination Engineering**—the science of managing collaborative teams.
|
||||
|
||||
### Key Breakthroughs:
|
||||
- **Portable Swarm Skills:** A new standard for sharing coordination protocols across different frameworks (e.g., moving a workflow from AutoGen to LangGraph) is emerging, allowing for "Zero-Adapter Deployment" of complex team behaviors.
|
||||
- **Self-Evolving Coordination:** Systems are now using **Trajectory Distillation** and **Friction-Driven Optimization** to automatically identify bottlenecks in agent communication and generate patches without human intervention.
|
||||
- **Extreme Specialization:** To minimize context-switching latency, the trend is moving toward "Elite Teams" where roles are decoupled into highly specific sub-roles (e.g., separating a "Financial Auditor" from a "Copywriter").
|
||||
|
||||
---
|
||||
|
||||
## Conclusions and Future Outlook
|
||||
The research of Q2 2026 suggests that the "Agentic" era is maturing. The focus is no longer on *if* an agent can perform a task, but on how reliably it can manage a **long-horizon trajectory** while maintaining a **persistent identity** and **collaborating within a swarm**. Future developments will likely center on "Routing Stability"—ensuring that as agents become more specialized, the handoffs between them remain seamless and error-free.
|
||||
|
||||
## References
|
||||
- [Do Agents Need to Plan Step-by-Step? Rethinking Planning Horizon](https://arxiv.org/html/2605.08477) (May 2026)
|
||||
- [DeepPlanning: Benchmarking Long-Horizon Agentic Planning](https://arxiv.org/html/2601.18137v1) (January 2026)
|
||||
- [Kimi K2 Thinking: Mastering Long-Horizon Agentic AI](https://i10x.ai/news/kimi-k2-thinking-agentic-ai-model) (2026)
|
||||
- [Goal2Skill: Long-Horizon Manipulation with Adaptive Planning](https://arxiv.org/html/2604.13942v1) (April 2026)
|
||||
- [Memory as Ontology: A Constitutional Memory Architecture](https://arxiv.org/html/2603.04740v1) (March 2026)
|
||||
- [Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification](https://arxiv.org/html/2605.10052v1) (May 2026)
|
||||
Reference in New Issue
Block a user