Skip to content
Research radar

What the research says, and what we do about it

A living list of the papers and reports that change how we design systems. Each entry is a primary source with the engineering decision we draw from it.

Research radar

Current radar

Updated as the field moves. Sources are linked; takeaways are ours.

every card links to the primary paper
AgentsAnthropic · 2024

Building effective agents

Five workflow patterns (chaining, routing, parallelisation, orchestrator-workers, evaluator-optimiser) cover most production needs. Add true agency only when the path cannot be known in advance. It is still the reference taxonomy, and the rule we design by.

EvaluationAnthropic · 2026

Demystifying evals for AI agents

Separate capability evals from regression evals, start with 20–50 seed tasks, calibrate LLM judges against humans, and gate releases. This is the shape of the evaluation loop we install in every system.

ReasoningDeepSeek (Nature) · 2025

DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

Reinforcement learning with rule-based rewards produced frontier-level reasoning in an open-weight model, which reset the cost of strong reasoning and widened the build-versus-buy options for narrow tasks.

CapabilityMETR · 2025–2026

Measuring AI ability to complete long tasks

The length of tasks agents complete at 50% reliability doubled roughly every seven months across 2019 to 2025, and on METR's updated suite the post-2024 doubling is closer to three months. Plan architectures for what models will do next year, but ship what they can do reliably today, with checkpoints.

RetrievalChroma · 2025

Context rot: how increasing input tokens impacts LLM performance

Across 18 frontier models, accuracy declined as inputs grew, well below the advertised context limits. Retrieval and context engineering are not obsolete; they are how long-context models stay accurate.

RetrievalACL 2026 Industry Track · 2026

Is agentic RAG worth it? An experimental comparison of RAG approaches

Agentic retrieval used 3.3× the input tokens and 1.5× the latency of a well-built enhanced pipeline, and the agent chose to retrieve again on only about one query in ten, where half the documents came back unchanged. Route simple questions to the fast path; reserve agents for the hard ones.

RetrievalMicrosoft Research · 2024–2025

LazyGraphRAG: a new standard for GraphRAG quality and cost

Graph-based retrieval answers whole-corpus questions vector search cannot, and the lazy variant indexes at roughly 0.1% of full GraphRAG's cost, which makes it practical for real corpora.

RetrievalAnthropic · 2024

Contextual retrieval

Prepending chunk-specific context before embedding, combined with BM25 and reranking, cut retrieval failures by 67%. A cheap, high-leverage upgrade we apply to every RAG system.

CostNVIDIA Research · 2025

Small language models are the future of agentic AI

Most calls inside an agent are narrow, repetitive subtasks that small models already handle. Routing those steps to small models is the largest single cost lever in agent systems.

CostLMSYS / UC Berkeley (ICLR 2025) · 2024

RouteLLM: learning to route LLMs with preference data

A learned router cut cost by over 85% while retaining 95% of frontier quality by sending only the hard minority of queries to the strong model. Routing is a design decision, not an afterthought.

ReasoningApple · 2025

The illusion of thinking

Reasoning models collapse beyond a complexity threshold and overthink easy problems. Route by difficulty rather than defaulting to the biggest model with the longest thinking budget.

ReliabilityOpenAI · 2025

Why language models hallucinate

Hallucination is partly a product of training and benchmarks that reward confident guessing over abstaining. We score abstention and calibration in our golden sets, not just correctness.

BenchmarksSierra · 2025

τ²-bench: evaluating conversational agents in a dual-control environment

Agents that follow policies while talking to users and calling tools are far less consistent across repeated runs than single-run scores suggest. We test consistency, not just accuracy.

DataXLang Lab (ICLR 2025) · 2025

Spider 2.0: evaluating language models on real-world enterprise text-to-SQL workflows

On real enterprise schemas that often exceed a thousand columns, the reference agent solved about 21% of tasks in the original result, and the leaderboard now passes 76% on Spider 2.0-Lite. Analytics copilots need a governed semantic layer under them, which is where our data work goes first.

SecurityOWASP GenAI Security Project · 2026

OWASP Top 10 for LLM applications and for agentic applications (2026)

The threat model for prompt injection, insecure tool use, excessive agency and data leakage. Our guardrail layers map to it item by item.

SafetyAnthropic / Redwood Research · 2024

Alignment faking in large language models

Models can behave differently when they believe they are being observed. For enterprise deployment this argues for least-privilege tools, confirmation on consequential actions and traces that are reviewed.

ProtocolsLinux Foundation AAIF · 2026

Model Context Protocol specification, 2026-07-28 revision

MCP won the agent-to-tool layer with SDKs from every major vendor; the 2026 revision simplifies sessions and deprecates older primitives. We expose systems to agents through MCP so they are portable across models.

IndustryGartner · 2025

Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027

Escalating costs, unclear value and inadequate risk controls kill agent projects. Our process starts with a metric, a baseline and a go/no-go gate precisely to avoid becoming that statistic.

IndustryMcKinsey / QuantumBlack · 2026

The state of AI in 2026: on the road to ROI

High performers are far more likely to redesign workflows around AI rather than bolt it on, and to put senior owners on the outcome, which is why discovery starts with the workflow and the metric.

IndustryStanford HAI · 2026

AI Index Report 2026

Open-weight models sit within a few percent of the closed frontier on public leaderboards, though the gap reopened in 2025 after nearly closing in 2024. Vendor-neutral architecture lets clients move with that gap rather than bet on one side of it.

The landscape we design for, with sources.

$12.5BEnterprise spend on foundation model APIs in 2025, out of $37B on generative AI overall (Menlo Ventures)
Source
40%+Of agentic AI projects Gartner expects to be cancelled by 2027 on cost, unclear value and weak risk controls
Source
85%Cost reduction from model routing on MT-Bench while keeping 95% of GPT-4 quality (RouteLLM, ICLR 2025)
Source
~4 monthsHow far open-weight models trail the closed frontier in Epoch AI's 2026 analysis, which keeps vendor-neutral designs cheap
Source

Want this applied to your problem?

A discovery call is where research meets your data.

Book a discovery callSend us a brief