AI in SWE: Enterprise Token Governance Meets Foundry at Scale
A 24-hour editorial briefing on Uber’s agentic coding caps, Microsoft Foundry’s Build follow-through, and the widening gap between agent capability and enterprise budgets.
Focused on items published or materially updated in the last 24 hours in Europe/London time.
AI in SWE: Enterprise token governance meets Foundry at scale
Wednesday, 3 June, was the day enterprise AI coding stopped sounding like a pilot and started sounding like procurement.
Uber’s spending caps dominated the narrative — WebProNews’s Wednesday analysis framed the $1,500 monthly limit per agentic tool as the end of the freewheeling experiment. Analytics India Magazine and Outlook Business added detail: limits apply only to agentic coding software such as Cursor and Claude Code, tracked separately per tool, visible on an internal dashboard, with manager-approved exceptions.
Meanwhile Microsoft published the Foundry at Build follow-through post — a build → deploy → operate story for agents at scale. Capability expansion and budget contraction in the same 24-hour window.
Why Uber’s cap is a benchmark, not an outlier
The numbers explain the policy. Reports cite 95% of engineers using AI tools monthly and a large share of committed code originating from agent-assisted workflows. Uber burned through its planned 2026 AI coding budget in the first four months. CTO Praveen Neppalli Naga’s “back to the drawing board” quote from earlier reporting now has a concrete instrument: per-tool token ceilings.
Outlook Business also notes Microsoft winding down internal Claude Code licenses in favor of GitHub Copilot CLI — a reminder that enterprise agent strategy is vendor portfolio management, not personal preference.
The important design choice is carving out agentic tools only. Uber still encourages other AI applications. Leadership is targeting the highest-burn category while preserving experimentation elsewhere.
Foundry’s answer: operationalize agents, don’t just license them
Microsoft’s Foundry blog describes production agents as a closed loop:
- Build: Agent Framework harness, skills in toolboxes, procedural memory, Voice Live — stay in familiar IDEs and frameworks.
- Deploy: Hosted agents, long-running routines, publishing to Teams and M365 Copilot.
- Operate: Tracing, evaluation, agent optimizer turning production failures into ranked improvements.
That is platform engineering for agents. Identity, sessions, observability, and rollout controls are the product — not the chat window.
The Agent Framework Build post adds MAF integration with the GitHub Copilot SDK: shell execution, file ops, MCP, streaming, OpenTelemetry — Copilot as a MAF backend agent.
Enterprises drowning in token bills may still need this layer. Governance without observability is just rationing without learning.
Two governance philosophies
Uber’s approach is entitlement governance: cap consumption at the source, per tool, with dashboards and exceptions.
Foundry’s approach is operational governance: host agents, trace failures, optimize prompts and tools from production signal.
You need both. Caps prevent runaway spend. Platforms prevent runaway behavior.
What I would do with this as an engineering leader
- Publish internal per-tool monthly budgets before engineers discover them via surprise throttling.
- Measure tokens per merged PR and per incident resolved — not tokens per developer curiosity.
- If you standardize on Foundry or similar, require traces for any agent with write access to repos.
- Negotiate exception paths that require task definition and expected outcome, not just “I need more tokens.”
- Compare Uber-style caps with model routing (cheaper defaults for routine agent steps).
Bottom line
June 3 crystallized the enterprise agent dilemma: adoption is already mainstream, budgets are not.
Uber’s caps are the first widely reported template for rationing agentic coding. Microsoft Foundry is the platform bet that agents become managed services with SLAs, not unchecked IDE experiments.
The next quarter’s winners will measure both — dollars and outcomes.