PLDI 2026

Basics

Field Value
Dates <2026-06-15 Mon><2026-06-19 Fri>
Venue Limelight Boulder, Boulder, Colorado
Host ACM SIGPLAN
URL pldi26.sigplan.org
Format Multi-track with co-located events, workshops, tutorials
Framing PL research, compilers, verification; PAgE workshop on agentic engineering

Co-located conferences

  • ISMM 2026 — International Symposium on Memory Management
  • LCTES 2026 — Languages, Compilers, and Tools for Embedded Systems

Workshops

Workshop Full name Notes
ARRAY 2026 Array-oriented Programming  
CP 2026 Choreographic Programming  
EGRAPHS Equality Graphs E-graph rewriting, optimization
PAgE 2026 Principles of Agentic Engineering Formal foundations for agent safety/reliability
PLMW PL Mentoring Workshop  
PROPL Programming for the Planet  
SOAP 2026 State of the Art in Program Analysis  
Tom Ball at Sixty Celebration session

Tutorials

  • Writing Performance-Portable Kernels Simplified with Helion
  • A guided tour through Oxidized OCaml
  • Implementing an Integrated Development Environment
  • Mechanized Specifications for Real-World Programming Languages
  • ACT: End-to-End Compiler Infrastructure for Emerging AI Accelerators
  • Deep dive into the AWS Nitro Isolation Engine

PAgE 2026 — Principles of Agentic Engineering

The workshop most directly connected to current research.

Field Value
Date <2026-06-15 Mon>
Deadline 2026-05-06 AoE
Portal page2026.hotcrp.com
Contact sbarke@microsoft.com

Organizing committee

  • Shraddha Barke (Microsoft Research, Redmond)
  • Peli de Halleux (Microsoft Research)
  • Dan Grossman (University of Washington)
  • Madan Musuvathi (Microsoft Research)
  • Dawn Song (UC Berkeley)
  • Ben Zorn (Microsoft)

Submission tracks

  1. Archived papers (up to 10 pages) — published in ACM Digital Library
  2. Non-archival submissions (up to 6 pages) — work in progress, position papers, demos

Why PAgE matters

"AI-powered agents increasingly deployed in production settings require formal foundations for safety and reliability through agentic engineering principles."

This is the PL community's formal-methods entry point into agent governance — the same territory BugBash approaches from the testing/DST side and ELS approaches from the Lisp/MOP side. The committee (Grossman, Song, Zorn) carries enough weight to set vocabulary that sticks.

Program (Mon 15 Jun 2026, Meadows CD)

Tentative; chaired by Shraddha Barke. Two keynotes bracket two sessions.

Time Talk Who
09:10 Keynote: Formal Methods for Frontier AI Systems Gagandeep Singh (UIUC)
10:40 Agentic Code Reasoning Ugare, Chandra (Meta)
11:00 Towards Verified Code Reasoning by LLMs Sistla et al. (Google DeepMind)
11:20 Testing LLM-Generated Distributed Protocol Code Das, Coyne (Boston U.)
11:40 The Next Frontier for AI-Generated Kernels: Correctness Martínez (MSR), Sorensen (UCSC)
12:00 Testing, Credible Compilation, and Verification in the Axon Verified Compiler in Lean and Claude Code Rinard (MIT)
13:40 Combining Agentic AI and Lightweight Formal Methods to Find Bugs in a Production Hypervisor Katsura, Memarian, Sewell (Cambridge)
14:00 From Workarounds to Root Causes: Agentic Workflows on Browser GPU Compiler Stacks Ramesh, Levine, Sorensen (UCSC)
14:20 Event-based Design Abstractions for Agent Harnesses Becker, Ghavami, … Rinard, Mansinghka (MIT)
14:40 Lumos: Let there be Language Model System Certification Chaudhary et al. (UIUC)
15:00 Lazy Validation and Self-Healing for Agentic Programs Tsampouris, Ioannidis (MSR), Symeonidis
15:50 Keynote: Can Coding Agents Write Verifiably Correct Software? Shan Lu (Microsoft / UChicago)

Closest to the reversible-pipeline verification thread: Rinard's Axon verified compiler in Lean and Claude Code and the Cambridge lightweight formal methods + agentic AI hypervisor work — both the verification-first, agent-drives-the-prover pattern we used for the Lean Equiv and Dafny codec proofs. Shan Lu's closing keynote — can coding agents write verifiably correct software? — is the experiment those proofs run.

Why

PLDI sits at the intersection of formal methods and practical compiler engineering. Three reasons this year matters:

  1. PAgE workshop. First SIGPLAN-affiliated venue explicitly framed as "agentic engineering." The formal-methods community claiming this term has downstream consequences for how agent governance gets specified. The elenctic-spec framing and Seven Concerns layering should be testable against whatever vocabulary PAgE settles on.
  2. Nitro Isolation Engine tutorial. AWS Nitro is the production hypervisor-isolation analog to Antithesis. The tutorial is a data point on how far hardware-backed isolation has moved since BugBash.
  3. Egraphs workshop. E-graph rewriting is the optimization substrate beneath equality saturation. Relevant to any spec-refinement system that needs to normalize equivalent representations — the same problem elenctic-spec faces at the methodology layer.

Refutation

The PAgE framing for this event is wrong if:

  • The workshop is dominated by LLM-agent benchmarking rather than formal foundations (safety/reliability becomes eval leaderboards).
  • The "agentic engineering" vocabulary diverges completely from the Seven Concerns framing, with no productive mapping between them.
  • The Nitro tutorial is marketing rather than technical substrate.

In those cases PLDI reduces to a compiler conference with an adjacent agent workshop that didn't converge on useful primitives.

Concern mapping

Event/Talk Concern Claim under test
PAgE workshop (overall) L3, L7 PL community can formalize agent governance primitives
Grossman/Song/Zorn committee L3 Formal methods vocabulary for agents will be PL-native
Nitro Isolation Engine L1, L5 Hardware isolation composes with agent sandboxing
Egraphs workshop L2, L3 Equality saturation applies to spec normalization
Mechanized Specifications L3 Mechanized specs are the reader-as-spec-emitter generalized
Oxidized OCaml tutorial L1, L2 Rust ownership model transfers to ML family

Related