Overview of 2025 was the year of agents, what's coming in 2026?
Practical AI hosts Daniel Whitenack (Prediction Guard) and Chris Benson (Lockheed Martin) reflect on major AI developments from 2025—particularly the rise of agentic AI—and forecast what to watch in 2026. The episode mixes concrete anecdotes (rapid developer productivity gains with agents), technical trends (reasoning models, multimodality), infrastructure and geopolitical implications (energy, chips), and practical advice for builders and organizations.
Key themes from 2025
-
Agents took center stage
- 2025 was the year discussion shifted from standalone models to autonomous agents (models + connectors + tooling) that complete goals across systems.
- Huge hype; real wins where teams had domain expertise and good integration, lots of failures where teams lacked that.
-
Reasoning models (the “reasoning era”)
- Newer models stream chain-of-thought-like tokens that mimic reasoning before an answer; helpful for complex, dynamic workflows.
- Tradeoffs: much higher latency and compute cost (each token is a model run), which can be undesirable for optimized, deterministic business workflows.
-
Multimodality: input > output
- Multimodal inputs (text + audio + images + video) are increasingly used in business.
- Multimodal outputs remain rarer in enterprise contexts—text/structured outputs still dominate; synthesized speech is an exception.
-
Generative vs. predictive models
- Generative models (transformer-based) appear to be plateauing in frontier gains; open-source models have largely caught up.
- Predictive/discriminative models (forecasting, classification) continue to improve rapidly and deliver strong ROI; these are often the workhorses in production.
-
Orchestration and tooling matter more than the single best model
- Value is coming from composing models, RAG systems, databases, and domain-specific tools behind orchestrators/agent controllers rather than owning a “best” model.
- AutoML/augmented ML approaches integrated as callable tools inside agentic systems are powerful.
-
Infrastructure, energy and geopolitics
- Bottlenecks are shifting from GPUs alone to power and energy availability; big implications for chip plants, power plants, and regional policy.
- National strategies and geopolitics are increasingly influenced by AI power/compute supply considerations.
What worked (anecdotes & evidence)
-
Real productivity jumps
- Daniel: company product development accelerated enough to impress investors without expanding headcount.
- Chris: a complex autonomy project—after many iterative prompts—produced in minutes what would have taken weeks, showing agentic tooling can be multiplicative.
-
Success factors
- Domain expertise + iterative prompting + correct data/tool connections.
- Good business-case selection (don’t automate broken processes).
What failed or remains hard
- Many agent projects fail without integration expertise.
- Reasoning-token latency and cost can make certain models impractical for high-throughput business flows.
- Multimodal output experiences and fully integrated physical-agent systems are still emerging and uneven.
Technical trends & implications
- Models
- Frontier generative model improvements slowing; open-source alternatives competitive.
- Reasoning/hybrid models are popular but costly; streaming tokens = more GPU runs.
- Tooling & architecture
- MCP servers, RAG, tool connectors, and orchestration layers are becoming essential skills.
- The valuable role emerging: engineers who can design and integrate multi-service agentic systems (AI integrators/AI engineers).
- Predictive stacks
- Forecasting and discriminative models remain critical—often wrapped as callable tools inside agentic orchestrations.
Infrastructure, hardware & geopolitical trends
- Energy is now as important as chip supply: datacenter power, reactivating decommissioned plants, local resistance to big fabs, and national strategy.
- Onshoring, chip plants, and energy availability feed geopolitical decisions—AI consumption is now a policy factor.
Predictions for 2026 (hosts’ forecasts)
- Continued rise of agentic systems where orchestration and tools, not just models, drive value.
- Growing demand for integration expertise: building MCP servers, RAG, secure connectors and orchestrations.
- The “AI maker”/consumer hardware era begins: cheaper embedded AI (GPUs/ASICs) enabling maker kits, consumer robotics/toys, and local experimentation.
- Fragmentation continues—ecosystem complexity remains high; consolidation will be gradual, creating opportunities for “quick time-to-value” vertical and compliance-focused solutions.
- Quantum computing: still not broadly practical in 2026 (hosts skeptical of near-term practical impact).
Practical recommendations / action items
- For builders and teams:
- Prioritize clear use cases and business value before agentizing a workflow.
- Invest in domain expertise + iterative prompt engineering; plan for many prompts to reach production quality.
- Architect for flexibility: avoid model lock-in; design systems that can use multiple back-end models and tools.
- Learn to build/connect MCP servers, RAG systems, and toolchains—this integration skill will be highly valuable.
- For leadership:
- Expect rapid feature velocity with small teams if agents and orchestration are implemented correctly—reassess hiring and productivity models.
- Prepare for compliance complexity (e.g., NIST standards); consider vendors that simplify secure, compliant deployments.
- For everyone:
- Be mindful of cost & latency of reasoning models—choose models/configurations that balance quality, speed, and power use.
Notable quotes / soundbites
- “Even I at moments am feeling a bit left behind with how fast this is changing.” — paraphrasing Andrej Karpathy on rapid tooling changes.
- “Agents are an alien tool with no manual—figure out how to hold it and operate it.” — hosts’ paraphrase of industry sentiment.
- “The model is no longer the blocking point—it’s the connectivity, tooling, and compliance that make systems hard.” — summary observation.
Resources & sponsors mentioned
- NIST AI standards (NIST SP 800-series / NIST 601 referenced for secure AI compliance)
- Framer (sponsor ad: framer.com/practicalai — promo mentioned)
- Prediction Guard (sponsor / Daniel’s company: predictionguard.com)
- Hosts/affiliations: Practical AI podcast (practicalai.fm), Daniel Whitenack (Prediction Guard), Chris Benson (Lockheed Martin)
Episode takeaway (one-paragraph)
2025 accelerated the move from models to agents: real upside when domain expertise, iterative prompting, and proper orchestration are combined, but many projects still fail without integration skills. Expect 2026 to emphasize orchestration, infrastructure and energy constraints, continued improvement in predictive models, and a growing market for engineers who can build end-to-end agentic systems—plus the first signs of a consumer “AI maker” hardware wave.