The three questions

Every UE5 developer who lands here is silently asking three things. The honest answers below trace back to a research scan run 2026-05-23 (sources catalogued under References).

Q1 — GOAP still relevant?

Yes, with a record that has strengthened in 2024–2025, not weakened.

The 2025 AI and Games Conference (London) — now the principal practitioner gathering for hardcore game AI — included three talks that bear on this directly:

• Eric Jacopin's AI Planning Analytics — From F.E.A.R. (2005) to Assassin's Creed: Shadows (2025) [14] is a twenty-year longitudinal study of planner analytics across two flagship shipped titles. The strongest single piece of evidence that GOAP-derived planning ships in current-generation AAA.
• Petr Smrček's Supporting Thousands of Simulated NPCs in the Open World of KCD2 [15] pairs with the existing Marko 2025 KCD2 talk [6] — the same studio published two production talks about the same shipped GOAP-derived system, one on architecture, one on scale.
• Jeff Orkin's One Trillion Parameters and No Plans [16] is the twenty-year retrospective from the creator of GOAP. He argues — directly addressing the elephant in the LLM room — that planners do things LLMs cannot, and that the future is neurosymbolic: LLMs streamline authoring, planners handle runtime. He demonstrates this with a Bitpart.AI prototype combining Claude and a planner.

If the originator of GOAP is publicly defending the paradigm two decades later in front of a room of practising AAA AI engineers, the case "this is a dead technique" does not hold up.

Q2 — Who it's for

Who it is NOT for: AAA studios shipping bespoke AI (they build in-house), linear action / shooter projects where StateTree or BTs already cover the design surface (Aversa estimates 80% of game AI needs [18]), crowd-scale simulations (Mass AI territory), conversation-only projects (Convai / Inworld), twitch combat / per-frame tactical scoring (Utility AI), and tightly-scripted boss choreography or cinematics (BT, StateTree, or Level Sequences).

If you read both lists and your project doesn't land on either, the Quickstart is the cheapest way to find out. Ten minutes; no commitment.

Q3 — The LLM era

LLM-NPC vendors (Convai, Inworld) sell conversation and expression: real-time voice, lip-sync, emotional response, character dialogue. They do not, and structurally cannot — sell behaviour runtime. Conversation is generated; behaviour at runtime needs to satisfy game rules, respect world state, replicate across the network, integrate with animation, and remain deterministic enough to test. Those are planner concerns, not language-model concerns.

The emerging consensus is neurosymbolic — LLMs at the authoring and dialogue layer, symbolic planners at the runtime execution layer. Orkin's 2025 talk argues this position directly [16]; his Bitpart.AI prototype demonstrates it in working code.

Intent Forge's forward-looking story is "the trusted runtime planner beneath your LLM authoring stack." No LLM tooling ships today; the scope is the runtime. But the architecture is intentionally LLM-receptive: archetype assets, actions, and goals are data the LLM can author. The dispatcher abstraction means an LLM-generated archetype slots into the same execution layer as a hand-authored one.

UE5 GOAP competitors

The shape of the table: one actively-maintained free alternative, one generic paid plugin, otherwise abandoned or narrowly-scoped projects. Intent Forge differentiates against all of them on the same three axes: stability (the anti-flap toolkit), networking (the server-authoritative replication design), and the dispatcher abstraction that composes with StateTree, BT, or external execution layers.

Unity reference point

The Unity ecosystem has a single dominant open-source GOAP plugin: crashkonijn/GOAP [8]. In March 2026 that project sits at v3.1.2, ~1.7k GitHub stars, 863 commits across 16 branches and 72 tagged releases, with Discord, GitBook documentation, OpenUPM packaging, and a Unity Asset Store presence. It originated as the production AI of the shipped game Basher Beatdown.

The UE5 ecosystem has the same demand and no comparable fully-productised plugin. The Unity scale of adoption is also useful as a sanity check on the addressable market: GOAP is not a niche-of-a-niche; it is a recognisable named technique with thousands of practitioners.

Epic's StateTree

StateTree is Epic's officially-supported, hierarchical state-machine-plus-selector substrate. It became the default UE AI authoring tool from UE 5.4–5.5 onward; community sentiment in 2025 reads "State Trees are the future, behaviour trees feel outdated."

StateTree is not a planner — behaviours are hand-authored, not synthesised. The framing question for any UE5 developer in 2026 is:

"When does the hand-authored substrate (StateTree) stop scaling, and a planner start to pay back?"

Aversa's canonical community essay [18] names the two motivating criteria: an action-space too combinatorial to enumerate ahead of time, or a context that cannot be fully anticipated at design time (emergent gameplay, user-generated content, sandbox/sim worlds).

The community-converged pattern is hybrid: StateTree at the top-level state machine, a planner inside the states where decisions actually emerge. Intent Forge's dispatcher abstraction is purpose-built for exactly that pattern — making it the complement to StateTree, not a competitor.

Mass AI

Mass AI is Epic's data-oriented framework for crowd simulation: thousands of agents with simple per-agent logic. Intent Forge targets a different agent count and complexity: tens to hundreds of expressive NPCs, not thousands of simple ones. A project building a city of background extras should reach for Mass AI; a project building a dozen agents that need to feel like they are thinking should reach for a planner.

LLM-NPC vendors

LLM-NPC vendors aren't competing with planner-based AI; they're competing with hand-authored dialogue trees. The two layers combine cleanly: an LLM handles what the character says and how it emotes; a planner handles what the character does next, why, and when. The neurosymbolic direction is a tailwind for Intent Forge, not a threat.

AAA shipping record

The 2024–2025 AAA record shows GOAP-derived planning shipping in:

• Kingdom Come: Deliverance II (Warhorse) — GOAP combined with Markov Behaviour Trees, with a separate published talk on scaling to thousands of simulated NPCs [6] [15].
• Assassin's Creed Shadows (Ubisoft) — twenty-year longitudinal analytics study covering planner behaviour from F.E.A.R. through this title [14].
• Earlier shipping evidence: F.E.A.R. (2005) [1], Tomb Raider (2013), Shadow of Mordor (2014), Assassin's Creed Odyssey / Immortals Fenyx Rising [4], Square Enix multi-agent cooperation [5].

AAA studios don't buy plugins for their core AI; they ship bespoke implementations. Intent Forge doesn't target that segment. The AAA evidence is here because it answers the prior question: GOAP is a live AAA technique in 2025, with multiple independent shipping titles each year.

Honest red flags

A positioning page that lists only strengths is propaganda. Three recurring criticisms of GOAP that prospective users will find on their own:

Verify it yourself

Everything above traces to public sources. To check the claims yourself in under an hour:

If those four sources lead you to a different conclusion, the project wants to hear it — please open an issue.