2025 marks a turning point as large language models embrace agentic capabilities and face new safety challenges

The year 2025 has seen a significant shift in large language models towards agent-like behaviour, with advances in reinforcement learning and deployment bringing new opportunities and safety concerns to the forefront of AI development.

The past year has hardened into a moment of reckoning for large language models (LLMs) and the organisations that build and depend on them. What began as a debate over whether LLMs were “stochastic parrots” has given way to a more tex...

tured consensus: these systems are neither mere mimicry nor ready-made minds, but rapidly evolving tools whose capacities and failure modes must be understood in context.

According to a curated week-in-review of AI developments, 2025 saw a string of technical and product advances that moved LLMs closer to agentic behaviour. Reinforcement learning techniques , written about widely under new labels such as Reinforcement Learning from Verifiable Rewards (RLVR) , have been deployed to embed planning, tool use and memory more tightly within model behaviour, producing better problem-solving and more sustained multi-step reasoning. Academic surveys of the year describe this shift as a move toward “model-native” agents, where models internalise not just translations of instruction but procedures for acting over time. Industry data and research papers cited in that corpus show reinforcement learning as the primary enabler of this transition.

Commercial advances in 2025 reinforced the perception of progress. OpenAI’s GPT-5, released in August 2025, and the subsequent GPT-5.2 family introduced in December 2025, formalised distinct operating modes , “instant” for retrieval- and latency-optimised behaviour and “thinking” for deeper chain-of-thought reasoning , and a Pro tier aimed at even stronger reasoning. According to the Wikipedia entries summarising those releases, the GPT-5 series expanded distribution through ChatGPT, Microsoft Copilot and the OpenAI API, while security assessments continued to flag potential misuse and prompt-injection risks. Those concerns were echoed in OpenAI’s public acknowledgement that AI browsers such as Atlas remain exposed to sophisticated prompt-injection attacks; the company described mitigations but conceded such attacks may never be eliminated entirely.

Google’s research push was another defining thread. Google Research framed 2025 as a year of translating lab advances , in multimodality, efficiency and factuality , into product and scientific impact. The company’s work on AI agents for cybersecurity, described in corporate briefings and technical notes, demonstrates how autonomous agents can triage and remediate routine threats, freeing human analysts to focus on complex incidents. Google’s approach to agentised security highlights both opportunity and risk: automation reduces toil but increases the need for robust guardrails against adversarial manipulation and systemic failure.

New architectures and the search for better inductive biases continued outside the hyperscalers. Pathway’s Dragon Hatchling, reported by Live Science, experiments with continuously adaptive internal connectivity modelled loosely on biological synaptic change; early tests placed it at parity with much older baselines on some benchmarks but signalled a possible direction for models that learn and generalise online rather than relying solely on static pretraining. Parallel to architecture work, an arXiv survey and other analyses emphasised that agentic capabilities can also arise from training and optimisation regimes rather than from radically different model wiring alone.

The field’s evaluation practices came under sustained scrutiny. Technical commentary in the week’s aggregation questioned overreliance on benchmark scores and small-sample inferences: papers and posts argued that benchmark outcomes depend heavily on harness settings, scoring choices and even seemingly minor configuration tweaks. The METR plot , an attempt to measure “horizon length” for models by task completion time , was criticised for thin sampling in longer-duration ranges, a limitation that can distort debates about timelines to more general capabilities. At the same time, new evaluation frameworks for context compression and agent memory reported that structured summarisation methods outperformed some proprietary approaches from major vendors in preserving useful context across long-running sessions, an important operational detail for persistent agents.

Operationalising agents in enterprises spurred further engineering work on tooling and deployment. Discussions of multiplexing MCP (model control plane) servers describe architectures that expose specialised toolsets to agents via gateways, trading complexity for richer, safer instrumented access. Such designs aim to let agents call specialised services without proliferating direct connections, an approach industry practitioners say improves security and governance while enabling task-specific performance.

Events during 2025 also served as reminders that even well-designed systems can fail in messy, human-facing ways. A San Francisco power outage left Waymo’s robotaxi fleet saturated by safety confirmation checks at darkened intersections, illustrating how layered safety measures can produce large-scale availability problems when multiple checks queue simultaneously. The incident is a cautionary analogue for AI deployments: safety logic that is correct in isolation can interact poorly at scale.

Broader societal and market trends reflected the technology’s spread and the tensions that follow it. TIME’s 2025 “Person of the Year” feature on the “Architects of AI” highlighted the centrality of hardware and platform companies , and the concentration of influence that comes with them , while reporting widespread concerns about misinformation, inequality and governance. Industry forecasts collected in the week’s review predicted that 2026 would see accelerated adoption of agents within businesses, changes to staffing and compensation strategies, and increased financial innovation around tokenised liquidity and stablecoins; such predictions underline the view that AI is as much an economic force as a technical one.

Security and robustness remained unresolved themes. OpenAI’s public notes on Atlas and coverage of automated red-teaming efforts show firms responding with layered defences including adversarial testing, RL-powered red teams and tighter runtime filters. Academic work on “model collapse,” analysing rises in semantic similarity across internet content, warned that recursive reuse of model outputs as training data could degrade data richness and model generalisation over time unless countermeasures are developed.

Hiring and human capital for ML teams also evolved. Practitioners and recruiters argued for interview processes that better reflect real-world work: coding assessments paired with data-modelling tasks and project deep-dives, rather than abstract puzzle-style interviews, are being recommended to identify engineers who can ship reliable systems in production.

Taken together, the reporting from this week portrays 2025 as a year when LLMs matured into tools with more coherent internal planning and memory, when agentic behaviours became a dominant engineering pattern and when deployment realities , from prompt-injection to compositional failure modes , forced a shift from exuberant proof-of-concept to cautious systems engineering. Progress is tangible: multimodal, lower-latency, reasoning-capable models now power consumer and enterprise applications at scale. Yet the work ahead remains substantial: better evaluation practices, improved datasets and protections against recursive contamination, and governance mechanisms that align capability deployment with societal priorities.

The dominant lesson of the year is pragmatic. Advances such as RLVR-style training, context-compression techniques and specialised control-plane architectures expand what agents can do, but they do not eliminate brittleness. As researchers and companies translate breakthroughs into products, the balance between capability and control will define whether 2026 deepens AI’s positive impact or amplifies its systemic risks.

Source: Noah Wire Services