From ca2114a8b5e61e80ed59725a8b1055551bdb05bf Mon Sep 17 00:00:00 2001 From: Federico Kamelhar Date: Tue, 27 Oct 2026 16:34:37 -0400 Subject: [PATCH] docs: drop the framework comparison matrix; keep orientation for LangGraph and Strands users The "How Locus compares" table (Locus vs LangGraph vs Strands Agents) is removed from the home page and the README, along with the README header link to it. docs/concepts/compare.md becomes "Coming from LangGraph or Strands": the mapping notes for people arriving from LangGraph, Strands or plain LangChain stay, the feature scoreboard about third-party projects goes. Nav label and the Why Locus link follow. Signed-off-by: Federico Kamelhar --- README.md | 17 ---------- docs/concepts/compare.md | 73 +++++++++++++++------------------------- docs/index.md | 14 -------- docs/why-locus.md | 2 +- mkdocs.yml | 2 +- 5 files changed, 29 insertions(+), 79 deletions(-) diff --git a/README.md b/README.md index 8bacfa55..e839366f 100644 --- a/README.md +++ b/README.md @@ -33,7 +33,6 @@ Multi-agent shapes · OCI GenAI models · The admission gate · - How Locus compares · Notebooks

@@ -194,22 +193,6 @@ full runtime. --- -## How Locus compares - -| | Locus | LangGraph | Strands Agents | -|---|---|---|---| -| Agent loop with tools | `Agent` + `@tool` | via LangChain | `Agent` + `@tool` | -| Graphs with state, reducers, interrupts | `StateGraph`, functional API | `StateGraph` | graph tool | -| Pipelines, orchestrator, swarm, handoff, debate | built in | compose yourself | graph, swarm, workflow | -| Cognitive router: task to protocol | built in | — | — | -| Admission gate and signed audit trail | `admit()`, `governed_agent()` | — | — | -| Grounding of every claim | GSAR | — | — | -| OCI Generative AI, DAC, Oracle Database 26ai | native | community packages | — | - -As documented by each project in October 2026. → [Full comparison](https://locusagents.ai/concepts/compare/) - ---- - ## Notebooks and deploy [`examples/`](examples/) holds 89 notebooks, numbered in reading order: start here (01–11), diff --git a/docs/concepts/compare.md b/docs/concepts/compare.md index 291e93b1..4e68282c 100644 --- a/docs/concepts/compare.md +++ b/docs/concepts/compare.md @@ -1,54 +1,35 @@ --- -title: How Locus compares -description: Locus next to LangGraph, LangChain and Strands Agents, feature by feature, as each project documents itself in October 2026. +title: Coming from LangGraph or Strands +description: "How Locus maps onto what you already know from LangGraph, LangChain and Strands Agents, and where it adds something." --- -# How Locus compares +# Coming from LangGraph or Strands Locus is an agent harness: the agent loop, the ways to combine agents, and the -runtime around them. The closest neighbours are LangGraph (with LangChain -underneath) and Strands Agents. This page lists where they meet and where -Locus differs, as each project documents itself in October 2026. Check the -other projects' docs for anything that matters to you; they move fast. - -## At a glance - -| Capability | Locus | LangGraph / LangChain | Strands Agents | -|---|---|---|---| -| Agent loop with tools | `Agent`, `@tool`, typed argument coercion, idempotent tools | agents and tools via LangChain | `Agent`, `@tool` | -| Graphs | `StateGraph` with reducers, conditional edges, `Send` fan-out, interrupts, checkpointed state; functional API | `StateGraph`, functional API | graph multi-agent tool | -| Composition | `SequentialPipeline`, `ParallelPipeline`, `LoopAgent` | compose from graph nodes | workflow tool | -| Orchestrator, swarm, handoff | built in, typed specialists and handoff context | supervisor and swarm libraries | swarm and graph tools | -| Debate, critic loops | notebooks and protocols | build yourself | build yourself | -| Cognitive router (task to protocol) | `Router.dispatch()`: typed `GoalFrame`, eight protocols, deterministic routing | — | — | -| Cross-process agents | A2A protocol client and server | via integrations | A2A support | -| Admission gate, signed audit trail | `admit()`, `approve()`, `governed_agent()`, hash-chained `AuditTrail` | human-in-the-loop interrupts | — | -| Grounding of claims | GSAR: every claim grounded, contradicted or abstained | — | — | -| Guardrails and steering | input, output and tool-call guardrails, LLM-as-judge steering, trace hygiene | via LangChain | guardrail hooks | -| Checkpointers | 10 backends: Oracle Database 26ai (plus versioned saver), OCI Object Storage, PostgreSQL, MySQL, Redis, OpenSearch, S3-compatible, memory, file, HTTP | PostgreSQL, SQLite, Redis, memory, others via packages | session managers | -| Long-term memory | `OracleStore`, Oracle agent memory, LLM memory manager | store API | via providers | -| RAG | six vector stores, OCI and in-database embeddings, Cohere rerankers, `retriever.as_tool()` | via LangChain | via tools | -| Observability | in-process `EventBus`, `run_context()`, 40+ typed events, SSE | LangSmith | OpenTelemetry | -| Serving | `locus serve` (FastAPI, bearer auth, threads), runner image | LangGraph Platform | AgentCore | -| OCI Generative AI | native: V1, SDK and Responses transports, auto-routing by model id | community packages | — | -| Oracle Database 26ai | native vector store, checkpointer, versioned saver, long-term memory, in-database chunking and embeddings | community packages | — | -| Other providers | OpenAI, Anthropic, 17 OpenAI-compatible routes (OpenRouter, Together, Groq, Ollama, vLLM, …) | broad | Bedrock-first, others via providers | - -## What this means in practice - -- **Coming from LangGraph.** The `StateGraph` surface will feel familiar: nodes, - edges, reducers, `Send`, interrupts, checkpointers. Locus adds the shapes you - would otherwise assemble from nodes (orchestrator, swarm, handoff, debate), a - router that picks the shape from the task, and a control layer that runs - outside the model. See [StateGraph](multi-agent/graph.md) and - [Compatibility](../compatibility.md) for the known API differences. -- **Coming from Strands Agents.** The agent loop, `@tool` and the multi-agent - tools map one to one. Locus is Oracle-first where Strands is AWS-first: OCI - Generative AI and Oracle Database 26ai are native transports and backends, - and the admission gate plus GSAR grounding have no counterpart. -- **Coming from plain LangChain.** Locus has no LangChain dependency. Its RAG, - memory and model layers are its own; the Oracle Database 26ai primitives in - particular do not go through `langchain-oracle`. +runtime around them. If you have built with LangGraph, LangChain or Strands +Agents, most of it will feel familiar. This page is the orientation. + +## Coming from LangGraph + +The `StateGraph` surface maps closely: nodes, edges, reducers, `Send` +fan-out, interrupts, checkpointers, and a functional API. Locus adds the +shapes you would otherwise assemble from nodes (orchestrator, swarm, handoff, +debate), a router that picks the shape from the task, and a control layer +that runs outside the model. See [StateGraph](multi-agent/graph.md) and +[Compatibility](../compatibility.md) for the known API differences. + +## Coming from Strands Agents + +The agent loop, `@tool` and the multi-agent tools map one to one. Locus is +Oracle-first: OCI Generative AI and Oracle Database 26ai are native +transports and backends, and the admission gate and GSAR grounding are part +of the runtime rather than add-ons. + +## Coming from plain LangChain + +Locus has no LangChain dependency. Its RAG, memory and model layers are its +own; the Oracle Database 26ai primitives in particular do not go through +`langchain-oracle`. ## Where to start diff --git a/docs/index.md b/docs/index.md index 996822f4..13d3a261 100644 --- a/docs/index.md +++ b/docs/index.md @@ -296,20 +296,6 @@ pip install locus-sdk && python -m locus.rogue [First controlled action →](how-to/first-controlled-action.md) · [Notebook 71 — Payment refund gate →](notebooks/notebook_71_payment_refund_gate.md) -## How Locus compares - -| | Locus | LangGraph | Strands Agents | -|---|---|---|---| -| Agent loop with tools | `Agent` + `@tool` | via LangChain | `Agent` + `@tool` | -| Graphs with state, reducers, interrupts | `StateGraph`, functional API | `StateGraph` | graph tool | -| Pipelines, orchestrator, swarm, handoff, debate | built in | compose yourself | graph, swarm, workflow | -| Cognitive router: task to protocol | built in | — | — | -| Admission gate and signed audit trail | `admit()`, `governed_agent()` | — | — | -| Grounding of every claim | GSAR | — | — | -| OCI Generative AI, DAC, Oracle Database 26ai | native | community packages | — | - -As documented by each project in October 2026. [Full comparison →](concepts/compare.md) - ## Where next [All notebooks →](notebooks/index.md) · diff --git a/docs/why-locus.md b/docs/why-locus.md index f6dafebc..8f2957a1 100644 --- a/docs/why-locus.md +++ b/docs/why-locus.md @@ -9,7 +9,7 @@ Locus is the Oracle Generative AI Agentic Harness: one `Agent` class, eight multi-agent shapes, and the runtime around the model (tools, memory, RAG, observability, guardrails), built first-class for OCI Generative AI and Oracle Database 26ai and open to every other provider. The [home page](index.md) -and [How Locus compares](concepts/compare.md) cover that ground. This page is +and [Coming from LangGraph or Strands](concepts/compare.md) cover that ground. This page is about the part that is easiest to over-claim: what the control layer guarantees, what you configure, and what stays model-dependent. diff --git a/mkdocs.yml b/mkdocs.yml index f8cf6dfd..c6bd2b6e 100644 --- a/mkdocs.yml +++ b/mkdocs.yml @@ -149,7 +149,7 @@ nav: - Agent loop: concepts/agent-loop.md - Tools: concepts/tools.md - Model providers: concepts/models.md - - How Locus compares: concepts/compare.md + - Coming from LangGraph or Strands: concepts/compare.md - Capabilities: capabilities.md - Multi-agent shapes: - Overview and the eight protocols: concepts/multi-agent.md