A season of agentic production: Fall 2026

Since mid-July this blog has published twenty pieces on putting agents into production: workflows before autonomy, evals, MCP, context, approvals, guardrails, cost, latency, retrieval, and injection. Mid-October is a good point to step back and ask which ideas held up, what changed during the season, and what we expect the next quarter to reward.

What held up

Short, supervised runs. The picture painted by Measuring Agents in Production in December 2025 has aged well. Its deployed systems mostly ran short trajectories under structured control flow, on off-the-shelf models, judged largely by human reviewers and built on custom harnesses, with reliability as the dominant concern. Nothing we have seen since suggests the successful pattern has shifted toward long, unsupervised autonomy.

The evaluation gap. LangChain's State of Agent Engineering reported 57% of respondents with agents in production, and observability far more widespread than evaluation. That gap is still the weakness we encounter most often in reviews. Teams can describe their runs in detail and still struggle to say whether a change made them better.

A shared security vocabulary. The OWASP Top 10 for Agentic Applications has become a practical common language. Saying "goal hijack" or "cascading failure" in a design review saves twenty minutes of explanation and points straight at specific controls.

What moved

MCP kept consolidating. The protocol's 2026-07-28 specification revision arrived early in the season, and employer demand kept rising. In Agentic Ready's September 2026 jobs index, MCP appeared in 588 postings across 2,331 open agentic roles at 193 companies. Our August argument, that MCP servers should be run like an API program with an inventory, scopes, schemas, and audit logs, feels more urgent now than when we made it.

The protocol layer widened. MCP is not the only standard in play. In April the Linux Foundation announced that the Agent2Agent (A2A) protocol had passed 150 supporting organizations, landed in major cloud platforms, and reached enterprise production use in its first year. The Agent Client Protocol (ACP), which connects coding agents to editors and other clients, counts Grok Build among the tools supporting it. Teams now have to reason about three kinds of connection: agent to tools, agent to agent, and agent to the interface where people work. Each needs its own identity model and its own audit trail.

What we underweighted

Two topics deserved attention earlier in the sequence than they got. The first is the human side of approvals: queue time, reviewer fatigue, and calibration matter as much as the model loop, yet they appeared only in the August piece on human-in-the-loop design. The second is trace data itself. Traces are the raw material for evaluation, but they also hold customer information, so retention and redaction rules need to be settled before the first production run rather than after the first audit question. If we were starting the season over, both would come in the first month, right after workflows.

What we expect the next quarter to reward

These are judgments, not forecasts backed by data:

1. Evidence over demos. Teams that can show an eval trend line, a trace of a blocked attack, and a rehearsed kill switch will win permission to expand autonomy. Teams that cannot will stay in pilot.
2. Boring integration. As protocols standardize, the advantage shifts from connecting systems to connecting them safely: scoped identities, approval tiers, and audit trails across all three connection types.
3. Operators, not just builders. Running agents day to day (reading traces, curating eval sets, tuning budgets) is becoming a recognizable role. Organizations that grow those people internally will move faster than those waiting to hire them.

The Agentic Production Series

These ideas form the backbone of the Agentic Production Series, BDArgentumAI's public training, running from October 27 to November 5, 2026. The four courses progress from building a first agent loop, to connecting MCP servers, tools, and memory, to multi-agent orchestration with contracts, to evaluation, guardrails, and deployment for production. Each is a full day at $750, roughly three-quarters hands-on, and participants keep the repositories they build. Dates and registration are on the events page at bdargentumai.com.

We also hold to a few things we will not do: no invented case studies, no fabricated testimonials, and no race-to-the-bottom positioning. The work should be judged on what it produces.

Into the winter

Over the coming months we plan to go deeper on three topics that kept surfacing this season: keeping eval sets current as products change, designing approval tiers for regulated workflows, and securing agent-to-agent connections. If one idea from these twenty posts is worth carrying forward, it is that autonomy should be earned with evidence, one rung at a time.

Sources cited

• Pan et al., Measuring Agents in Production (arXiv 2512.04123, submitted December 2025)
• LangChain, State of Agent Engineering (survey of 1,340 respondents, November 18–December 2, 2025; report published February 2026)
• Model Context Protocol specification, revision 2026-07-28 (released July 28, 2026)
• Agentic Ready (Prefactor), Agentic AI Jobs Index — September 2026 report (snapshot September 28, 2026)
• Linux Foundation, “A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms, and Sees Enterprise Production Use in First Year” (April 9, 2026)
• x.ai, Introducing Grok Build (May 25, 2026), which lists Agent Client Protocol (ACP) support
• OWASP GenAI Security Project, Top 10 for Agentic Applications 2026 (released December 9, 2025)
• BDArgentumAI, Agentic Production Series course pages at bdargentumai.com/events.html (as published in October 2026)