🌏 中文版
This is the final article. The previous 14 articles went deep: landscape, training, architecture, tools, evaluation, applications.
This one covers the future. Not predictions, but possible directions derived from current trends.
What's Been Done vs. What Hasn't
Let's be honest about the current state:
| Already Achieved | Not Yet Achieved |
|---|---|
| Search the web and synthesize reports | Propose original hypotheses |
| Cross-source cross-verification | Design and execute experiments autonomously |
| Produce citations-backed reports | Submit papers and pass peer review |
| Keep searching until "satisfied" | Keep improving until "correct" |
Current deep research agents are powerful assistants, but not yet autonomous researchers.
Direction 1: Self-Evolving Agents
From Tools to Agents
AREX already demonstrated the "dual-loop self-improvement" prototype. The next generation is:
Self-evolving agents: Not just correcting their own errors, but modifying their own strategies, architectures, and even learning methods.
From search results:
- Self-Evolving Agents survey (Gao et al., 2025, 277 citations): Established the framework for "what, when, how, and where to evolve"
- AlphaEvolve: Coding agents discovering new algorithms
- AutoResearchClaw: Automates the entire scientific lifecycle, from hypothesis to NeurIPS-ready PDF
- FARS (Analemma AI): Ran for 417 hours, produced 166 AI-generated papers
Key Questions
Where is the boundary of safe self-evolution?
- Safe evolution: Modify strategies while preserving core values
- Unsafe evolution: Change behavior to "perform better," potentially producing unintended consequences
- Human oversight's role: At what point does evolution require human approval?
Direction 2: Swarm Collaboration
From Single Agent to Agent Swarms
Claude Code's deep research already triggered 199 parallel sub-agents (accidentally). The future is intentional swarms:
- MiroThinker: Multi-model collaboration, BrowseComp 75.3
- MiroFlow: Top-1 on 5+ benchmarks, supports multiple models
- SWARMRESEARCH: Orchestrating coding agents for open-ended discovery
- MiroFish: Swarm intelligence engine under $1
Swarm vs. Single Agent
| Swarm | Single Agent | |
|---|---|---|
| Breadth | Better (multi-agent parallel exploration) | Limited |
| Depth | May scatter | Better (focused path) |
| Cost | High | Low |
| Consistency | Hard to maintain | High |
| Best for | Multi-angle tasks | Deep-dive tasks |
An Uncomfortable Reality
Gartner predicts: over 40% of agent projects will be cancelled by end of 2027. Usually because "the swarm was pointed at the wrong task."
Not that the technology doesn't work—it's that when to use a swarm and when not to isn't yet well understood.
Direction 3: Scientific Automation
From Research Assistant to Scientific Infrastructure
Ultimately, the ultimate form of deep research agents isn't "helping you do research" but becoming part of the scientific infrastructure itself:
Hypothesis → Experiment Design → Automated Execution → Data Analysis → Paper Writing → SubmissionEvery step handled by an agent, humans only make final judgments.
What's Already Happening
- FARS: 417 hours → 166 papers
- AutoResearchClaw: Full scientific lifecycle automation
- FAROS (OpenNSWM-Lab): Blueprint-driven AutoResearch runtime
- AlphaEvolve: Discovering new algorithms
Key Challenges
- Reproducibility: Can AI-generated experiments be reproduced?
- Honesty: Will AI fabricate data?
- Attribution: Who is the author of AI-generated papers?
- Quality control: Who reviews what AI reviews?
Convergence of Three Directions
Self-evolving × Swarm × Scientific Automation = Autonomous Scientific Research Infrastructure
Self-evolving agent swarm → Automated scientific process → Hypothesis to paper, fully autonomousBut this raises a fundamental question:
When AI can do scientific research autonomously, what is the human's role?
Not "replaced"—but "upgraded": from doing research to asking questions, setting directions, judging value.
The Uncomfortable Truth
One final point, cross-verified from multiple sources:
- Gartner: 40% of agent projects cancelled by 2027
- Enterprise rollback rate: 74% of enterprises have rolled back production AI agents
- Technology vs. application: Technology advances faster than application maturity
- Skill fragmentation: 10+ deep-research skills means methodology hasn't converged
The technology is ready. Humans aren't ready for how to use it.
What It Means for Us
This series of 16 articles itself does one thing: helping readers understand the field to make better judgments.
The ultimate goal isn't for readers to "know every detail"—but for them to:
- Know what choices are available
- Know the trade-offs of each choice
- Know when to use what
- Know where things might be headed
References
- A Survey of Self-Evolving Agents — Gao et al., 277 citations.
- AgentSwarm Playbook 2026 — Swarm application map.
- AlphaEvolve — Coding agent scientific discovery.
- FARS (Analemma AI) — 417 hours, 166 papers.
- AREX: Towards a Recursively Self-Improving Agent — Order 3 of this series, dual-loop prototype.
- deep-research-survey-overview — Series starting point: three-phase landscape classification.
Loading...