The Enterprise Agent Deployment Question Needs a Better Answer Before We Can Call It
textak holds this forecast at 77%, but we need to be honest about something: 'widely deployed' is doing too much work as a resolution criterion. Today's GitHub infrastructure data — 275 million commits per week, 17 million AI agent-opened pull requests in March 2026, Microsoft routing traffic through AWS because agents overwhelmed the platform — is the most direct deployment-scale evidence we've seen. It is also almost entirely about one domain: software development. Before we argue the thesis, we need to define what we're actually forecasting.
Let's start with the resolution problem, because it's real and the editorial flags are right to surface it. Our current forecast target — 'autonomous agents widely deployed in enterprise workflows' — cannot be independently resolved. A reader could look at GitHub's 14 billion annualized commits and $1 billion+ in AI coding tool revenue and argue this already resolved YES. Another reader could look at 88.4% platform availability and zero documented regulated-industry deployments and argue it hasn't. Both positions are defensible against the current target language, which means the target is broken.
Here is how we're tightening the resolution criterion going forward: the forecast resolves YES when agents are confirmed in production workflows — not pilots, not internal tools — at 20 or more Fortune 500 companies across at least two distinct enterprise functions (coding counts as one; customer service, finance, legal, or supply chain as others), with documented uptime SLAs or equivalent contractual commitments. This operationalizes what 'widely deployed' actually means and prevents the coding-only evidence from single-handedly resolving a forecast that was always meant to span enterprise functions broadly.
With that said, let's be honest about what today's evidence actually proves and what it doesn't. The GitHub numbers are extraordinary and represent direct evidence of production-scale agent deployment in software development workflows. These are not pilots. GitHub processed 275 million AI-generated commits per week in April — and we should note that 'AI-generated' here includes both autonomous agent commits and copilot-assisted commits, so the autonomous agent signal is likely a subset of that headline figure, with the AI agent-opened pull request count (4 million in September 2025 to 17 million by March 2026) being the cleaner metric for autonomous agent activity specifically. Even conservatively interpreted, the PR trajectory is not an experimentation number. The Cursor acquisition at $60 billion — reached two years after founding — is an investment signal, not a deployment confirmation, but it reflects market consensus that this revenue is real and durable, not speculative.
Here is what today's evidence does not prove: that agents are deployed at scale in customer service, financial workflows, legal operations, or supply chain — the other enterprise functions our original thesis cited. The lede mentioned customer service alongside coding, but all the quantitative evidence is coding-specific. We are currently supporting a multi-function enterprise deployment thesis with single-function evidence. That asymmetry matters. We're watching it.
The infrastructure reliability problem deserves a direct answer rather than a sidestep. Enterprise production systems require 99.9% availability — three nines — as a baseline SLA. GitHub's 88.4% availability during peak agent load in May 2026, which caused nine documented service incidents, represents roughly 1,050 hours of degraded service annually at that rate. That is not enterprise-grade. The fact that Microsoft is temporarily routing GitHub through AWS to manage the load is both evidence that agent deployment is real enough to break infrastructure AND evidence that the reliability foundation for enterprise-grade deployment hasn't been built yet. We haven't found documented cases of enterprises rolling back agent deployments because of the May incidents — and that absence is actually modestly positive for the thesis, suggesting enterprises are absorbing the reliability issues rather than retreating — but the absence of documented rollbacks is not the same as confirmed durability.
The 77% probability reflects this mixed picture. The positive evidence — GitHub production scale, PR trajectory, infrastructure investment signals — adds roughly four points from our previous assessment. The infrastructure reliability risk and the evidence asymmetry (coding-heavy, no regulated industry data) subtract roughly three points. Net: one point upward from 76% to 77%. We're not moving more than that because the strongest element of our thesis — that this is happening broadly across enterprise functions — remains supported by proximate evidence (cloud providers shipping agent frameworks, efficiency gains in pilots) rather than direct evidence of multi-function Fortune 500 production deployment. What would push us to 85%: Q3 earnings calls from two or more major banks, insurers, or manufacturers confirming agent deployments with operational metrics. What would drop us below 65%: documented enterprise rollbacks attributed to reliability failures, or two consecutive quarters of flat PR growth on GitHub suggesting the coding wave has plateaued without adjacent function adoption following.