Apollo GraphQL Summit 2026 - San Francisco

Overview

Apollo Summit 2026 ran October 6–8, 2026, at The Midway, 900 Marin St, San Francisco. Event page and agenda.

The livestream is the Wednesday keynote only: Apollo Summit 2026 Keynote Livestream, 1:07:55, started 08:46 PT on October 7. The agenda lists the keynote at 09:00 on the main stage, with Matt DeBergalis (CEO). Times below are video offsets.

Stream coverage

None of the Wednesday sessions after the keynote are in the video. The stream ends with the keynote's close; the breakout talks, lightning talks and workshops were not streamed here.

Agenda item (Wed Oct 7, PT) Speakers Video
09:00 Apollo Keynote Matt DeBergalis, guests 0:00 (captions from 18:37)
09:30 Fireside chat: where agentic systems are headed Angie Jones, Matt DeBergalis, Jeff Auriemma not on stream
10:30 Design, Govern, Automate: PayPal's Playbook Sarath Siva, Mike Hamrah not on stream
10:30 Workshop: Using GraphOS Agent Services Michelle Mabuyo not on stream
11:00 Agents to Rule Them All: a Context Graph at RS Group Juan Carlos Blanco Delgado not on stream
11:00 Lightning: What's New in Schema Proposals Zack Warnimont not on stream
11:25 Lightning: How Wiz Generates Its Terraform Provider Hen Perez, Efraim Globus not on stream
11:30 What Survives? Durable Architectures (Starbucks) Sharon Gorla not on stream
11:50 Lightning: Wayfair's Query Plan Pre-Warming Harsh Pai not on stream
13:00 Intuit's Persisted Queries: A Safety Layer for Agents Alan Cai not on stream
13:00 Workshop: Evolving Your Graph Without Breaking It Michelle Mabuyo not on stream
13:30 Trust First: American Airlines' Graph Safe for Agents Punj Prakash, Surya Saputra not on stream
14:00 Your Graph Is a Model of Your Business (Block) Jonathan Wondrusch not on stream
14:30 Doubling Down on the Core: Router 3 Speedups Jonathan Rainer, Taylor Ninesling not on stream; previewed at 1:01:30

On stage the Router 3 talk is announced for "Thursday at 2:30"; the agenda above lists it on Wednesday. The keynote also says a full Agent Services session with a demo follows it; which agenda slot that is was not stated.

Keynote segments

Segment Speaker(s) Video
APIs as the context layer; semantics and rules of use Matt DeBergalis 18:37
The domain model: semantics, policy, guidance Matt DeBergalis 25:00
GraphOS Agent Services announced Matt DeBergalis 28:00
Field-level access policy Matt DeBergalis 30:00
Intuit marketing "brain" (video, then on stage) Chris Miller (Intuit), Stephanie Wiseman (10KR) 32:00
Building the domain model: GraphOS Factory Matt DeBergalis 43:30
GraphOS MCP server, Rover, GraphOS Desktop preview Matt DeBergalis 47:00
Agent experience: one graph, opened to agents platform engineering lead, "Enable" 49:30
Operations and quality Josh Burgle (head of engineering) 54:00
Router 3 preview; layered pipeline Josh Burgle 58:00
Circuit breaking, rate limiting, new query planner Josh Burgle 1:00:30
Control plane and client updates Josh Burgle 1:04:00
Recap and close Matt DeBergalis 1:06:30

Top segments

Picked for concrete numbers and for overlap with research here: agents and MCP, contract governance, trust boundaries, query planning.

GraphOS Agent Services: policy on the fields, not the tools

28:00

  • Four services: search, agent identity, access policy, audit. They run in a new layer above the GraphOS routers and are exposed as MCP tools.
  • The layer merges one federated supergraph or many separately built graphs into one catalog, the "domain model".
  • Policy is defined centrally on domain-model fields and checked per request against the requesting agent. Examples given: a support agent sees the customer name but not rewards; a marketing agent sees rewards but not the name; a desktop analyst agent can request escalation, approved by a person.
  • A rule on a field holds whether the agent asks for it directly, reaches it through an order object, or writes a query nobody anticipated. No per-tool policy.
  • The domain model is split into semantics, policy (deterministic, enforced, audited), and guidance (when to make the expensive AS400 inventory call). Available on the day of the keynote.

Intuit and 10KR: from 19 local MCPs to governed agents

34:30

  • Goal stated as real-time reallocation of paid search spend; one campaign had a $60,000 cost per acquisition, and a recommendation sat for weeks.
  • MVP built in Claude Cowork against about 19 custom MCP servers, read and write, with no identity management and single points of failure on individual laptops.
  • Framed as the BFF-sprawl problem recurring as agent sprawl. The three gaps named: on whose behalf the agent acts, what it may see and do, and what happened afterwards (identity, policy, audit).
  • Reported results: double-digit improvements in cost per acquisition and incremental return on ad spend; pilot done in under 90 days. No baseline given.

GraphOS Factory, MCP server and Rover

45:00

  • Factory takes API docs or an OpenAPI/Swagger spec, walks a modeler through design decisions, and emits an Apollo Connectors subgraph for the router.
  • Each decision is recorded with its sources in git, so a later API change or a new API is checked against earlier choices, and hand-offs between modelers keep the same rationale. Ships as a skill in Apollo's skills collection on GitHub.
  • The GraphOS MCP server now covers most of GraphOS and Studio, writes included: building graphs, schema checks, graph health, proposals. Listed in the ChatGPT and Claude marketplaces.
  • Rover gains contract management, client operation checks and an expanded plugin system. GraphOS Desktop shown as an unreleased preview.

Opening one graph to agents ("Enable")

50:00

The captions render the company as "Enable": cybersecurity and IT management, millions of managed devices, grown by acquisition. That matches N-able; the name was not checked.

  • Acquisitions left systems that disagreed on what a device was; a backup-status question could hit several APIs. "Project Fusion" unified them into one graph API.
  • After Apollo's MCP server shipped, the whole API was opened to agents in about two months, with no rebuild, because the graph was already designed and documented.
  • A prospect's own agent ran a comparison against a competitor and recommended them; the deal was won. "Your number one API client isn't a developer anymore. It's an agent."

Router 3 and the beam-search query planner

58:00

  • First preview release of Router 3. The request pipeline is refactored into layers with single responsibilities; in Router 2, a change to metrics touched many services. The stated aim is the tail (P99 and above), not the median, and smaller change surfaces for agent-written changes.
  • New query planner adapted from beam search in the literature. Across a corpus of the most complex customer graphs: 20x faster planning and 20x less memory. Worst operation went from 28 s to 90 ms.
  • Circuit breaking for struggling subgraphs, to avoid congestive collapse, lands for GA; distributed rate limiting follows. Native Rust plugin interface changes before GA, target end of year; Apollo asked for production feedback since it lacks customers' extensions and configs.
  • Operations context: incidents down 34% year over year; support reproduces 9 of 10 reported issues before escalation and ships 11x more direct fixes; 2 trillion operations a month through GraphOS.
  • Also listed: graph artifacts for rollbacks, canaries and blue-green publishing; Swift 6 support with a 75% speed-up the captions render as "decalization", likely deserialization, on iOS.