Loading...
Loading...
Data
The browsable, downloadable data behind our research. Free to use with attribution (CC BY 4.0).
Reference路111 harnesses路updated 2026-10-25
The list of AI coding agents we track across our research. Each entry gives the maker, the type of tool, whether it is open source, and official links.
Catalogue路117 services路updated 2026-10-25
The building blocks developers use to make AI agents: places to run them, model access, sandboxes, data and memory, monitoring, testing, and safety tools.
Catalogue路61 evals路updated 2026-10-25
Tests that measure how well AI agents do real tasks. For each test: what it checks and how it marks answers. Where it applies, we also note whether the best agents have nearly maxed it out.
Catalogue路96 models路updated 2026-10-25
Current speech-to-speech, text-to-speech, speech-to-text, voice conversion, music and sound-effect models, with access, price, latency, voice-cloning policy, licence and status.
Catalogue路24 models路updated 2026-10-25
Models that return a typed decision with probabilities instead of text: general decision models, rerankers, classifiers, guardrails, scorers and reward models, with access, licence, price and what each source claims about calibration.
Catalogue路84 regulations路updated 2026-10-25
AI laws and rules for people who build with AI, in the US and the EU. Each entry shows who it covers, whether it is in force, key dates, and what it requires.
Catalogue路70 tools路updated 2026-10-25
Tools on both sides of the contest over AI-made content: detectors, watermarks, proof-of-origin standards, quality scorers, and tools that hide or remove AI signs.
Catalogue路8 figures路updated 2026-10-24
Published AI spend figures harmonised onto one denominator where that is defensible and kept apart where it is not, each row carrying its population, scope, basis, statistic and period.
Catalogue路12 programmes路updated 2026-10-17
Named programmes that run software production like a factory line, from AI agent platforms to government software units, with claims kept apart from evidence.
Catalogue路14 studies路updated 2026-10-17
Studies of how AI changes the speed and quality of software work. Each entry shows who was studied, what was measured, the result, and who paid for the study.
Matrix路62 primitives路updated 2026-10-16
A side-by-side comparison of AI coding assistants and the core features each one offers, such as plan mode, helper agents, hooks, memory, and restore points.
Catalogue路4 observations路updated 2026-10-16
Published studies of AI agents that post on public sites and react to posts by other agents. Figures come from other researchers, never from our own crawl.
Catalogue路29 models路updated 2026-10-14
The best open AI models to run on your own computer or phone. Sorted by memory size, from phones and 8 GB laptops to 512 GB workstations.
Catalogue路80 pipelines路updated 2026-10-13
Step-by-step recipes that chain AI models to make images, video, audio, and 3D. Examples are draft to final, image to video, and consistent characters.
Catalogue路49 patterns路updated 2026-10-12
Reusable ways to set up AI agents as a team, such as a manager with helpers or agents that check each other. Grouped by type and maturity.
Catalogue路24 protocols路updated 2026-10-12
The shared standards that let AI agents use tools, talk to each other, work in apps, and make payments. Covers MCP, A2A, x402 and more, with history and plans.
Catalogue路94 agents路updated 2026-10-12
AI agent products that finish multi-step tasks for people on their own, from personal assistants to agents for business and specific industries.
Catalogue路13 mechanisms路updated 2026-10-12
Ways AI agents began to work together that no one designed or approved. Each entry gives the signs that reveal it and the planned method it resembles.
Catalogue路11 substrates路updated 2026-10-12
Types of public web space that AI agents have used to leave hidden messages. Lists kinds of site, never live addresses, with warning signs and fixes.
Catalogue路28 vector classes路updated 2026-10-12
The ways an AI agent can break out of its sandbox, send data out, or gain extra access. Each entry gives the control that failed and the fix.
Catalogue路6 cases路updated 2026-10-12
Cases where something an AI system made or learned carried over into a later session, release, or training run, and the control that stops it.
Catalogue路6 subjects路updated 2026-10-12
AI labs and sandbox providers scored on eight safety controls for the test environments where agents run. Every score comes from their own public documents.
Catalogue路34 tools路updated 2026-10-12
Tools that use an AI coding agent to do one job on existing code. Examples are code review, security checks, test writing, and code upgrades.
Matrix路46 technique classes路updated 2026-10-12
The stages of an attack by AI agents, from first look to final damage. Each stage lists known technique types, matched to the MITRE ATT&CK framework where possible.
Catalogue路36 provisions路updated 2026-10-10
Laws and company rules on copying a voice or face with AI, such as the ELVIS Act and NO FAKES. Shows what each one covers and its status.
Catalogue路34 storefronts路updated 2026-10-10
Shops and services that let AI agents pay per use through x402, a web payment standard. Every entry is checked against a primary source.
Catalogue路49 breakthroughs路updated 2026-10-10
Research breakthroughs made with help from AI. Each entry records how the result was made, how much people helped, and how scientists received it.
Catalogue路31 incidents路updated 2026-10-10
Real security incidents where an AI agent, not a person, took the harmful actions. Each entry gives the timeline, the type of weakness used, and how it came to light.
Catalogue路20 roles路updated 2026-09-25
New jobs for people who put AI to work inside organisations, rather than build it. For each role: what it does, the skills it needs, and typical pay.
Catalogue路22 topologies路updated 2026-09-22
How one error in a team of AI agents spreads to the rest. Each entry shows the trigger, how far the damage goes, how to spot it, and how to stop it.
Matrix路48 cells路updated 2026-07-12
Our tests of AI agent teams against a single model on maths and code tasks. Shows success rates, token costs, and how well each team stops a mistake.