Strategic foresight · Agentic engineering
A forward-looking map of where agentic engineering and AI are heading. Fifteen signals, plotted by how soon they matter and how ready you should be to act. The future fans out to the right — the further out, the wider the range of plausible outcomes. Click any signal to read its story.
Write the specification first, let agents implement against it. Tooling like Microsoft’s Spec Kit formalises the pattern — spec as the contract, code as the output. It shifts engineering effort from writing code to writing intent precisely.
The file taxonomy around agents is settling — CLAUDE.md for operating rules, AGENTS.md for team definitions, memory and context files, SKILL.md for workflows. Teams that treat these as maintained assets outperform teams that treat them as setup chores. Our module 03 / 06 / 19 territory becoming industry standard.
The successor discipline to prompt engineering — deciding what information reaches the model, when, and in what form. Selection, compression, isolation. The skill that separates working agents from flaky ones.
Model Context Protocol has crossed into mainstream adoption (tens of millions of downloads) — becoming the USB of agent-to-tool connection. Every internal tool without an MCP server is invisible to agents.
Agents operating real desktops and browsers as a new automation layer, replacing brittle RPA. Our module 10 territory. The constraint is reliability and permissioning, not capability.
Coordination is the new scale frontier — teams of specialised agents with task graphs, mailboxes, and isolated workspaces. Our module 20 covers the mechanics.
The emerging job title for what most “AI engineers” actually do — building the environment around the model (tools, knowledge, permissions, observation) rather than the intelligence itself. Expect role definitions and hiring to follow the term.
Static benchmarks are saturating. The frontier is interactive evaluation — agents judged on multi-step tasks in live environments (ARC-AGI-3 style). For buyers, demo performance stops being proof; task-level evals become procurement criteria.
Adoption is outpacing control. Agents need identities, scoped credentials, audit trails, and revocation — same as human employees. The first serious agent-caused incidents will make this a board topic.
Gemma, Phi, Qwen-class models running cheap, local, and private. The pattern: big models for reasoning, small models for high-volume narrow tasks. It changes the cost equation for production agent fleets.
Enterprise architecture where many agents discover, trust, and transact with each other across the org. The successor question to “which agent do we buy”: how do our agents interoperate.
Agent memory graduating from session hack to infrastructure layer — shared, queryable, governed. Whoever owns the memory layer owns the switching costs.
Agents that improve from their own deployment experience instead of waiting for the next base model. Trajectory data becomes a strategic asset — today’s logs are tomorrow’s training signal.
Models that learn how environments behave, not just how text continues. If they mature, agents gain planning and physical-world competence that LLMs lack. Watch, don’t bet yet.
Compliance moving from policy documents into the runtime — policy-as-code gating what agents can do, with the EU AI Act as forcing function. Governance becomes an engineering deliverable, which is good news for whoever builds it early.