Two threads run through the history of AI...
From Algorithms to Agents: Two Papers Reshaping Modern AI
Two threads run through the history of AI: the classical — elegant algorithms that solve problems with mathematical guarantees — and the emergent — systems that learn, plan, and act in ways their creators did not fully anticipate. In 2025, both threads produced remarkable research. I dug into two papers that sit at opposite ends of this spectrum, and together they tell a surprisingly coherent story about the present and future of intelligent systems.
"What makes agentic AI stand out is its combination of strategic planning, memory that preserves context over time, external tools to extend capabilities, and collaboration with other agents — all in one system."
— Bandi et al., 2025
Paper One: The Rise of Agentic AI
Title: The Rise of Agentic AI: A Review of Definitions, Frameworks, Architectures, Applications, Evaluation Metrics, and Challenges
Authors: Bandi et al.
Published: Future Internet, MDPI — September 4, 2025
DOI: 10.3390/fi17090404
This is a systematic review of 143 primary studies spanning 2005–2025. It covers LLM-based and non-LLM agentic systems, architectural taxonomies, evaluation metrics, and open challenges.
What Is Agentic AI, Exactly?
The term "AI agent" has been thrown around loosely for years. One of the most valuable contributions of this paper is bringing terminological rigour to a chaotic landscape. The authors position agentic AI as fundamentally distinct from both generative AI (which responds) and traditional automation (which follows fixed rules). An agentic system is defined by its ability to perceive, plan, act, and reflect — autonomously pursuing a goal across multiple steps and time horizons.
Over 90% of the 143 papers surveyed were published in 2024 or 2025. The field is not just growing — it is erupting.
The Five Pillars of an Agent
| Pillar | What It Means |
|---|---|
| Planning | Decompose complex goals into sub-tasks; adapt when steps fail |
| Memory | Short-term context + long-term knowledge retrieval across sessions |
| Tool Use | Call APIs, search the web, write code, interact with environments |
| Reflection | Self-critique outputs and iteratively refine toward the goal |
| Collaboration | Coordinate with specialized sub-agents in multi-agent pipelines |
Architectures: From Single Agent to Swarm
The paper classifies agentic architectures along a spectrum. At one end is the single-agent loop — an LLM repeatedly sampling actions until a goal is satisfied. At the other end are multi-agent systems (MAS) where orchestrator agents delegate tasks to specialist subagents, each with its own tools and memory. Hybrid architectures — partly symbolic, partly neural — are gaining ground for tasks demanding verifiable reasoning.
Key finding: More than 90% of surveyed agentic AI papers were published in 2024–2025, confirming that agentic AI is not a mature field with settled answers — it is an ongoing, high-velocity research frontier. Frameworks and best practices are still in flux.
Where Agents Are Being Deployed
The paper catalogs real-world applications across sectors:
- Education — adaptive tutoring, automated grading
- Healthcare — clinical decision support, patient monitoring
- Transportation — autonomous navigation, logistics
- Software Engineering — code generation, testing pipelines
- Customer Service — conversational agents handling complex queries
Each domain surfaces domain-specific challenges — hallucination is catastrophic in medicine, latency kills user experience in customer service.
The Open Challenges
| Challenge | Why It Matters |
|---|---|
| Explainability | Multi-step agent chains become black boxes. Tracing why an agent took a harmful action is hard. |
| Generalization | Agents trained for one domain often fail surprisingly in adjacent ones. |
| Bias & Transparency | Agents amplify biases encoded in LLMs; audit trails are sparse. |
| Job Displacement | Agentic systems can now fully automate knowledge-worker tasks previously thought safe. |
| Evaluation | No consensus benchmark exists for comparing agentic systems fairly across task types. |
My Insight
What struck me most is the evaluation gap. In classical ML we have GLUE, ImageNet, BLEU — standardised yardsticks. In agentic AI, two systems claiming "state-of-the-art" may have never been tested on the same benchmark. Until the community converges on rigorous, reproducible evaluation suites, progress claims will remain hard to interpret. This isn't a minor inconvenience — it's a structural obstacle to safe deployment.
Paper Two: A Smarter A* with Adaptive Weights
Title: Research on the A* Algorithm Based on Adaptive Weights and Heuristic Reward Values
Published: World Electric Vehicle Journal, MDPI — March 4, 2025
DOI: 10.3390/wevj16030144
This paper proposes a modified A* algorithm combining diagonal-free five-way search, adaptive heuristic weighting, and obstacle-corner safety filtering to improve path quality and runtime for autonomous vehicle navigation.
The Classic A* and Why It Struggles
The A* algorithm is one of the most elegant ideas in computer science. By combining the actual cost from start to the current node (g(n)) with an estimated cost to the goal (h(n)), it finds shortest paths efficiently — in theory. In practice, traditional A* faces four stubborn problems:
- Long runtimes on large grids
- Excessive node expansions
- Tortuous zig-zag paths
- Corner collision — where a path grazes the sharp corner of an obstacle, safe on a grid but disastrous for a physical robot
The classic formula is:
f(n) = g(n) + h(n)
The paper's adaptive modification:
f(n) = g(n) + w(n) · h(n) + R(n)
where:
w(n) = adaptive weight (adjusts based on proximity & obstacle density)
R(n) = heuristic reward value (incentivises goal-directed moves)
The Three Innovations
1. Diagonal-Free Five-Way Search
Traditional A* uses 4-directional (cardinal) or 8-directional (including diagonals) neighbourhood search. The paper replaces this with a five-way search that selects directions based on coordinate changes, tilting exploration toward the goal without wasting effort on diagonal neighbours near obstacles. The result: fewer nodes expanded with no loss in path quality.
2. Adaptive Weight w(n)
Rather than using a fixed heuristic weight throughout the search, the algorithm dynamically adjusts w(n). When the agent is far from obstacles, it boosts the heuristic weight to accelerate goal-directed expansion. When near obstacles, it reduces the weight to prioritise safety. This mirrors how a cautious driver speeds on an open highway but slows in a crowded parking lot.
3. Heuristic Reward Values R(n)
The reward value acts as a positive incentive — nodes that make genuine progress toward the goal receive a bonus in the evaluation function. Crucially, when the search neighbourhood contains obstacles on diagonal sides, diagonal moves are filtered out entirely, preventing corner-clipping and the physical collisions it causes.
Results: The improved algorithm achieves significantly fewer expanded nodes and shorter, collision-safe path lengths compared to standard A* — validated in environments directly applicable to intelligent electric vehicle navigation.
My Insight
What fascinated me here is the philosophy of adaptive weighting. Classical heuristic search treats the weight as a fixed knob — you tune it once and commit. This paper shows that making the weight context-sensitive — a function of the local environment — unlocks significant gains without sacrificing guarantees. It's a microcosm of a broader shift in AI: moving from rigid, pre-specified rules toward systems that adapt their own parameters on the fly.
Connecting the Dots
At first glance these two papers seem to live in different worlds: one discusses emergent LLM-based agents; the other refines a 1968 graph-search algorithm. But look closer and the same animating idea surfaces in both.
Both papers are fundamentally about one thing: making AI systems contextually intelligent — responsive to the specifics of their situation rather than blindly applying a fixed strategy.
In the A* paper, the heuristic weight adapts to obstacle density. In the agentic AI paper, agents reflect on their outputs and adjust their plans. In both cases, good performance requires sensing the local context and updating behaviour accordingly.
There is also a striking convergence in application domain. The improved A* targets autonomous vehicle navigation. Agentic AI systems are being deployed in transportation logistics. In a near future where a self-driving fleet is managed by an orchestrating AI agent that uses A*-style planners as sub-routines for real-time navigation, these two research threads literally run on the same hardware, toward the same goal.
What This Means for AI Practitioners
Classical algorithms like A* are not being replaced by LLMs — they are being integrated into agent pipelines as reliable, interpretable tools. When an agentic system needs guaranteed shortest-path computation, it will call a pathfinding module. That module may well be a variant of what this 2025 paper proposes. Understanding both ends of the AI spectrum — the learned and the engineered — is what will make you a well-rounded practitioner.
Final Thoughts
Reading these two papers back-to-back left me with a clearer mental model of the AI landscape. The agentic AI paper revealed how rapidly the field is racing toward autonomous, goal-driven systems — and how immature the evaluation and safety infrastructure still is. The A* paper reminded me that deep, focused engineering on a specific problem still produces clean, impactful results.
If I had to distil one lesson from both: adaptivity is the common currency of modern AI. Whether you're weighting a heuristic function or designing a multi-agent orchestration system, the systems that will win are the ones that sense their context and update accordingly.
The gap between classroom theory and deployed systems is closing faster than most curricula can track. Papers like these are the on-ramp.
References
Bandi, A. et al. "The Rise of Agentic AI: A Review of Definitions, Frameworks, Architectures, Applications, Evaluation Metrics, and Challenges." Future Internet 17(9):404, 2025. https://www.mdpi.com/1999-5903/17/9/404
"Research on the A Algorithm Based on Adaptive Weights and Heuristic Reward Values."* World Electric Vehicle Journal 16(3):144, 2025. https://www.mdpi.com/2032-6653/16/3/144
Top comments (0)