With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Core Problem and Mission
Ditto is designed to solve the problem of fragmented, short-lived context in AI conversations. Most AI assistants operate statelessly, meaning they forget previous user interactions. Ditto’s mission is to give these agents a persistent ‘second brain’ by maintaining a shared memory. As explained on the official site, Ditto acts as a “shared place to remember, retrieve, and act” for AI agents, effectively turning ongoing conversations and connected tools into durable context. This means Ditto aims to ensure that past user preferences, tasks, and conversation history are retained and organized, rather than lost between chat sessions.
Bittensor Subnet Operation
On Bittensor, each subnet defines how miners create its output and validators score it. Ditto operates as a Bittensor subnet (netuid 118) under this model. Miners on any subnet run inference or compute tasks and publish their results, while validators compare and score those results. In Ditto’s case, registered nodes (miners and validators) would be dedicated to the memory task defined by Ditto’s incentive rules. In general, validators in a subnet send queries to miners’ Axon endpoints and evaluate their responses, then allocate TAO rewards based on those scores. Specifically, the Bittensor documentation notes that “validators send requests to [miners’] Axon and evaluate responses. This drives the incentive mechanism, awarding emissions to the miner”. While Ditto’s exact query format isn’t publicly detailed, it would follow this pattern: miners host Ditto’s memory database or model and answer memory-related queries, and validators judge the relevance or accuracy of those answers according to Ditto’s own criteria. By successfully retrieving the right bits of contextual knowledge or making correct contextual inferences, a miner would earn higher scores and thus more TAO rewards.
Miner and Validator Roles
In practice, a Ditto miner’s contribution would involve producing the memory-based outputs that the subnet values. For example, a miner might take an input query (such as a question about past conversation topics) and use Ditto’s semantic memory graph to generate an answer or recall relevant information. Validators, on the other hand, would score those outputs. They might compare a miner’s answer against expected context or check the coherence of retrieved memories. The validators’ scores then update trust metrics (VTrust) and ultimately determine how the block emissions are split among miners. Although the technical details (e.g. model types or training datasets) are not public, this is the standard miner/validator loop in a Bittensor subnet: miners produce AI-generated content according to the subnet’s purpose, and validators vote on quality to reward the best contributions.
Final Output and User Experience
The end product of Ditto is its collective memory that agents can query. In user-facing terms, Ditto operates as a kind of decentralized knowledge graph. Users or developers interacting with Ditto (for instance through an API or integrated app) would experience an AI assistant that “remembers” prior details. For example, Ditto automatically clusters conversations into “Subjects” and visualizes how ideas connect over time. It thus provides semantic recall: one could ask about anything mentioned months ago, and Ditto retrieves the context. In effect, the subnet’s output is the ongoing alignment of its neural “neurons” around a common memory state. If similar memory-DB projects are any guide, one might think of Ditto as continuously indexing embeddings of past dialogues and serving them on demand. A related project, Engram (on Bittensor testnet), describes itself as a “decentralized vector database” for AI, storing embeddings and serving queries. Ditto’s mainnet output would play a similar role, but optimized around personal conversational context and real-world tasks as defined by Omni Aura’s team.
Comparison to Other Subnets
Ditto’s focus on AI memory makes it quite distinct among Bittensor’s subnets. Many other subnets target different AI problems. For example, the “Data Universe” subnet (SN13) is built to collect and share large training datasets across sources. Others like vision or roleplay subnets train or serve generative models. By contrast, Ditto’s niche is context and memory for an individual’s agents. One parallel is Project Nobi’s “Nori” agent, which also bills itself as a personal AI companion that remembers the user’s life, though Nori runs on a Bittensor testnet. In general, Ditto appears to be among the first mainnet subnets explicitly dedicated to personal AI memory. Its decentralized memory graph could complement subnets focused on data or content (like Engram or Data Universe) by providing the “personal knowledge” layer. This sets Ditto apart as a specialized substrate enabling context persistence, rather than general compute or data indexing.
Core Problem and Mission
Ditto is designed to solve the problem of fragmented, short-lived context in AI conversations. Most AI assistants operate statelessly, meaning they forget previous user interactions. Ditto’s mission is to give these agents a persistent ‘second brain’ by maintaining a shared memory. As explained on the official site, Ditto acts as a “shared place to remember, retrieve, and act” for AI agents, effectively turning ongoing conversations and connected tools into durable context. This means Ditto aims to ensure that past user preferences, tasks, and conversation history are retained and organized, rather than lost between chat sessions.
Bittensor Subnet Operation
On Bittensor, each subnet defines how miners create its output and validators score it. Ditto operates as a Bittensor subnet (netuid 118) under this model. Miners on any subnet run inference or compute tasks and publish their results, while validators compare and score those results. In Ditto’s case, registered nodes (miners and validators) would be dedicated to the memory task defined by Ditto’s incentive rules. In general, validators in a subnet send queries to miners’ Axon endpoints and evaluate their responses, then allocate TAO rewards based on those scores. Specifically, the Bittensor documentation notes that “validators send requests to [miners’] Axon and evaluate responses. This drives the incentive mechanism, awarding emissions to the miner”. While Ditto’s exact query format isn’t publicly detailed, it would follow this pattern: miners host Ditto’s memory database or model and answer memory-related queries, and validators judge the relevance or accuracy of those answers according to Ditto’s own criteria. By successfully retrieving the right bits of contextual knowledge or making correct contextual inferences, a miner would earn higher scores and thus more TAO rewards.
Miner and Validator Roles
In practice, a Ditto miner’s contribution would involve producing the memory-based outputs that the subnet values. For example, a miner might take an input query (such as a question about past conversation topics) and use Ditto’s semantic memory graph to generate an answer or recall relevant information. Validators, on the other hand, would score those outputs. They might compare a miner’s answer against expected context or check the coherence of retrieved memories. The validators’ scores then update trust metrics (VTrust) and ultimately determine how the block emissions are split among miners. Although the technical details (e.g. model types or training datasets) are not public, this is the standard miner/validator loop in a Bittensor subnet: miners produce AI-generated content according to the subnet’s purpose, and validators vote on quality to reward the best contributions.
Final Output and User Experience
The end product of Ditto is its collective memory that agents can query. In user-facing terms, Ditto operates as a kind of decentralized knowledge graph. Users or developers interacting with Ditto (for instance through an API or integrated app) would experience an AI assistant that “remembers” prior details. For example, Ditto automatically clusters conversations into “Subjects” and visualizes how ideas connect over time. It thus provides semantic recall: one could ask about anything mentioned months ago, and Ditto retrieves the context. In effect, the subnet’s output is the ongoing alignment of its neural “neurons” around a common memory state. If similar memory-DB projects are any guide, one might think of Ditto as continuously indexing embeddings of past dialogues and serving them on demand. A related project, Engram (on Bittensor testnet), describes itself as a “decentralized vector database” for AI, storing embeddings and serving queries. Ditto’s mainnet output would play a similar role, but optimized around personal conversational context and real-world tasks as defined by Omni Aura’s team.
Comparison to Other Subnets
Ditto’s focus on AI memory makes it quite distinct among Bittensor’s subnets. Many other subnets target different AI problems. For example, the “Data Universe” subnet (SN13) is built to collect and share large training datasets across sources. Others like vision or roleplay subnets train or serve generative models. By contrast, Ditto’s niche is context and memory for an individual’s agents. One parallel is Project Nobi’s “Nori” agent, which also bills itself as a personal AI companion that remembers the user’s life, though Nori runs on a Bittensor testnet. In general, Ditto appears to be among the first mainnet subnets explicitly dedicated to personal AI memory. Its decentralized memory graph could complement subnets focused on data or content (like Engram or Data Universe) by providing the “personal knowledge” layer. This sets Ditto apart as a specialized substrate enabling context persistence, rather than general compute or data indexing.
Current Status
As of now, Ditto’s core product is live in a limited sense. The Ditto memory assistant (hosted at heyditto.ai) is available to users – the site reports “760+ people” and “37,500+ conversations remembered”, indicating an active user base. On the blockchain side, Ditto has minted its own subnet token (the SN118 “alpha” token), which is tradeable on TAO-based markets. For instance, CoinGecko shows the SN118 token trading around $2.10, with a market capitalization of roughly $3.9 million. This implies Ditto has at least reached the stage where its token is circulating and miners can stake and mine. A Bittensor block explorer shows that SN118 is currently at full 256 UIDs with 14 validators and only 1 active miner, suggesting the subnet is just starting and has room to onboard more participants. In summary, the initial layers – user-facing memory interface and the subnet infrastructure on-chain – are up, while future development (scaling the subnet, attracting more miners, etc.) appears to be ongoing.
Technical Architecture
Ditto’s underlying architecture centers on AI language models combined with smart memory storage. According to the Ditto blog, the system uses a multi-layer memory approach (short-term, long-term episodic, and a semantic knowledge graph). This likely means Ditto maintains a database of text embeddings or graph links to previous conversations. The GitHub repositories hint at the tech stack: for example, the ‘nlp_server’ repo (part of the ditto-assistant organization) is described as hosting intent, NER models, and “an LLM agent with long term memory vector store” for Ditto clients. Similarly, a ‘vision_server’ repository holds models for image captioning and Q&A, used by an image retrieval LLM agent. These suggest Ditto’s platform uses neural networks (likely in Python/PyTorch) with custom memory management. The mention of “Memory Chain Protocol” (MCP) integration on their site confirms there is a component – possibly an API server – to serve memory queries to other tools. Overlaying all this is a web-based UI (the Ditto chat interface and subject dashboard) that ties together user inputs and memory outputs. Although the exact model sizes or database engines are not public, the presence of vector memory stores indicates use of embedding-based search. The architecture appears to be modular: separate services for NLP, vision, and memory management, likely orchestrated through containers (the “ditto-stack” repo suggests Docker-based local deployment capabilities).
Repository and Code Activity
The official code for Ditto’s Bittensor subnet is not publicly visible, but the developer’s GitHub organization (ditto-assistant) offers insight into related tooling. This org contains multiple projects: an NLP server, a vision server, a mobile app, and other utilities. Activity is modest: key repos like ‘nlp_server’ and ‘vision_server’ show updates in 2024 (June 2024). The ditto-assistant org has a few dozen followers and stars across its repos, suggesting a small core team. There is no separate public repo labeled for SN118, implying the Bittensor-specific code may be private or embedded within these projects. In terms of network metrics, TaoPulse indicates SN118’s daily emission share was around 1.13% of the total leaderboard. The alpha token supply (inferred from the $3.9M cap and $2.10 price) appears to be only a few million units, making it relatively scarce. No detailed on-chain mining stats (like total staked TAO) are published yet. Overall, the technical build shows typical use of deep learning and vector DB approaches, and the development activity level is consistent with a small startup team iterating on the core memory technology.
Integrations and APIs
Ditto is built to integrate with other AI tools. The website advertises a “Memory Chain Protocol” (MCP) with server and client components, indicating an API that any AI agent can use to access Ditto’s memory. It also lists integrations such as Google Workspace – e.g. allowing Ditto to read email, calendar, and documents as context during chat sessions. In practice, a developer could use Ditto’s MCP interface to connect Ditto’s memory graph to a custom chatbot or LLM: the agent would send user context to Ditto’s server and receive relevant memory data in return. On the developer side, the ‘ditto-stack’ repository shows they provide a way to run the full Ditto system locally, which could be used for testing or custom integrations. However, we found no publicly documented SDK beyond the GitHub code and mention of MCP – any detailed developer guide is not available. In summary, Ditto seems designed to plug into the broader AI ecosystem via web APIs and integrations (like MCP and Google services), but external developers will rely on (limited) documentation and code samples from the team.
Current Status
As of now, Ditto’s core product is live in a limited sense. The Ditto memory assistant (hosted at heyditto.ai) is available to users – the site reports “760+ people” and “37,500+ conversations remembered”, indicating an active user base. On the blockchain side, Ditto has minted its own subnet token (the SN118 “alpha” token), which is tradeable on TAO-based markets. For instance, CoinGecko shows the SN118 token trading around $2.10, with a market capitalization of roughly $3.9 million. This implies Ditto has at least reached the stage where its token is circulating and miners can stake and mine. A Bittensor block explorer shows that SN118 is currently at full 256 UIDs with 14 validators and only 1 active miner, suggesting the subnet is just starting and has room to onboard more participants. In summary, the initial layers – user-facing memory interface and the subnet infrastructure on-chain – are up, while future development (scaling the subnet, attracting more miners, etc.) appears to be ongoing.
Technical Architecture
Ditto’s underlying architecture centers on AI language models combined with smart memory storage. According to the Ditto blog, the system uses a multi-layer memory approach (short-term, long-term episodic, and a semantic knowledge graph). This likely means Ditto maintains a database of text embeddings or graph links to previous conversations. The GitHub repositories hint at the tech stack: for example, the ‘nlp_server’ repo (part of the ditto-assistant organization) is described as hosting intent, NER models, and “an LLM agent with long term memory vector store” for Ditto clients. Similarly, a ‘vision_server’ repository holds models for image captioning and Q&A, used by an image retrieval LLM agent. These suggest Ditto’s platform uses neural networks (likely in Python/PyTorch) with custom memory management. The mention of “Memory Chain Protocol” (MCP) integration on their site confirms there is a component – possibly an API server – to serve memory queries to other tools. Overlaying all this is a web-based UI (the Ditto chat interface and subject dashboard) that ties together user inputs and memory outputs. Although the exact model sizes or database engines are not public, the presence of vector memory stores indicates use of embedding-based search. The architecture appears to be modular: separate services for NLP, vision, and memory management, likely orchestrated through containers (the “ditto-stack” repo suggests Docker-based local deployment capabilities).
Repository and Code Activity
The official code for Ditto’s Bittensor subnet is not publicly visible, but the developer’s GitHub organization (ditto-assistant) offers insight into related tooling. This org contains multiple projects: an NLP server, a vision server, a mobile app, and other utilities. Activity is modest: key repos like ‘nlp_server’ and ‘vision_server’ show updates in 2024 (June 2024). The ditto-assistant org has a few dozen followers and stars across its repos, suggesting a small core team. There is no separate public repo labeled for SN118, implying the Bittensor-specific code may be private or embedded within these projects. In terms of network metrics, TaoPulse indicates SN118’s daily emission share was around 1.13% of the total leaderboard. The alpha token supply (inferred from the $3.9M cap and $2.10 price) appears to be only a few million units, making it relatively scarce. No detailed on-chain mining stats (like total staked TAO) are published yet. Overall, the technical build shows typical use of deep learning and vector DB approaches, and the development activity level is consistent with a small startup team iterating on the core memory technology.
Integrations and APIs
Ditto is built to integrate with other AI tools. The website advertises a “Memory Chain Protocol” (MCP) with server and client components, indicating an API that any AI agent can use to access Ditto’s memory. It also lists integrations such as Google Workspace – e.g. allowing Ditto to read email, calendar, and documents as context during chat sessions. In practice, a developer could use Ditto’s MCP interface to connect Ditto’s memory graph to a custom chatbot or LLM: the agent would send user context to Ditto’s server and receive relevant memory data in return. On the developer side, the ‘ditto-stack’ repository shows they provide a way to run the full Ditto system locally, which could be used for testing or custom integrations. However, we found no publicly documented SDK beyond the GitHub code and mention of MCP – any detailed developer guide is not available. In summary, Ditto seems designed to plug into the broader AI ecosystem via web APIs and integrations (like MCP and Google services), but external developers will rely on (limited) documentation and code samples from the team.
Development Team
Publicly available information identifies the Ditto project with an AI startup called “Omni Aura,” as shown on the official site and the GitHub profile. Omni Aura appears to be the creator and is framed as “built by the Omni Aura team.” Specific team members are not disclosed by name. The GitHub page does list a contact email “[email protected]”, implying at least one developer (Omar) is involved, but details beyond that are scarce. Omni Aura’s LinkedIn mentions that they “started building Hey Ditto before ChatGPT launched,” emphasizing the concept of persistent AI memory, but it does not list team bios. There is no indication of outside funding or formal partnerships in public sources. For community outreach, Ditto has a minimal social presence: for example, the @heydittoai Twitter account had only a few dozen followers as of late 2024. No other major backers or collaborations (e.g. with universities or companies) are mentioned. In the broader Bittensor ecosystem, other personal-AI-memory projects (like Project Nobi’s Nori) exist, but these are separate efforts. To summarize, Ditto’s team is primarily the unnamed Omni Aura developers (one of whom is likely Omar based on the GitHub), focusing internally on the product. The only public-facing indicators are the site copy and a few social posts; no additional team or advisor info has been published.
Development Team
Publicly available information identifies the Ditto project with an AI startup called “Omni Aura,” as shown on the official site and the GitHub profile. Omni Aura appears to be the creator and is framed as “built by the Omni Aura team.” Specific team members are not disclosed by name. The GitHub page does list a contact email “[email protected]”, implying at least one developer (Omar) is involved, but details beyond that are scarce. Omni Aura’s LinkedIn mentions that they “started building Hey Ditto before ChatGPT launched,” emphasizing the concept of persistent AI memory, but it does not list team bios. There is no indication of outside funding or formal partnerships in public sources. For community outreach, Ditto has a minimal social presence: for example, the @heydittoai Twitter account had only a few dozen followers as of late 2024. No other major backers or collaborations (e.g. with universities or companies) are mentioned. In the broader Bittensor ecosystem, other personal-AI-memory projects (like Project Nobi’s Nori) exist, but these are separate efforts. To summarize, Ditto’s team is primarily the unnamed Omni Aura developers (one of whom is likely Omar based on the GitHub), focusing internally on the product. The only public-facing indicators are the site copy and a few social posts; no additional team or advisor info has been published.
Future Plans and Vision
As of now, there is no formally published roadmap or milestone list for Ditto. The team has not announced specific phases or timelines on public channels. (For comparison, unrelated subnets have laid out multi-phase plans in their documentation, but no analogous information has emerged for Ditto.) Omni Aura’s communications emphasize the importance of AI memory – their LinkedIn post states they wanted persistent context “from day one” – but provide no schedule. The only hint of progress is a recent social media announcement: the team tweeted that “Live mode is finally here!”, suggesting the core service has entered general availability. Beyond this, anything in 2026 is speculative. A fully-realized vision of Ditto would presumably be a seamlessly integrated memory system for AI agents: an always-on knowledge graph and natural-language interface across all tools and workflows. However, no concrete targets (such as new language support, model upgrades, or on-chain integrations) are described in available sources. The team’s focus appears to remain on refining Ditto’s memory-assistant functionality. In summary, without any public roadmap announcements, we can only infer Ditto’s direction from its stated mission: keep building out the “second brain” concept for AI. Future development is likely to enhance the memory capabilities and integrations, but exact plans have not been shared.
Future Plans and Vision
As of now, there is no formally published roadmap or milestone list for Ditto. The team has not announced specific phases or timelines on public channels. (For comparison, unrelated subnets have laid out multi-phase plans in their documentation, but no analogous information has emerged for Ditto.) Omni Aura’s communications emphasize the importance of AI memory – their LinkedIn post states they wanted persistent context “from day one” – but provide no schedule. The only hint of progress is a recent social media announcement: the team tweeted that “Live mode is finally here!”, suggesting the core service has entered general availability. Beyond this, anything in 2026 is speculative. A fully-realized vision of Ditto would presumably be a seamlessly integrated memory system for AI agents: an always-on knowledge graph and natural-language interface across all tools and workflows. However, no concrete targets (such as new language support, model upgrades, or on-chain integrations) are described in available sources. The team’s focus appears to remain on refining Ditto’s memory-assistant functionality. In summary, without any public roadmap announcements, we can only infer Ditto’s direction from its stated mission: keep building out the “second brain” concept for AI. Future development is likely to enhance the memory capabilities and integrations, but exact plans have not been shared.
A special thanks to Mark Jeffrey for his amazing Hash Rate series! In this series, he provides valuable insights into Bittensor Subnets and the world of decentralized AI. Be sure to check out the full series on his YouTube channel for more expert analysis and deep dives.
Recorded in June 2026, Mark Jeffrey speaks with Payton of Ditto (Subnet 118), an agentic operating system built around persistent memory and the idea that future software should understand and remember its users across applications, agents and devices. Payton explains how Ditto’s memory system goes beyond basic vector search by building a structured knowledge graph of topics, relationships and user intent, allowing it to recall information with much greater accuracy and context while also acting as a shared “brain” for external agents such as Claude, OpenClaw and Hermes. The conversation covers Ditto’s Artifacts system for collaborative documents and apps, full data portability from services such as ChatGPT and Claude, and the upcoming Ditto Admin platform that will let developers build applications directly on top of users’ knowledge graphs. Payton also outlines plans for one-click Bittensor miners combined with live subnet knowledge graphs, helping agents stay current with constantly changing mining strategies and documentation, as well as Ditto Code, a beginner-friendly coding environment that combines an open-source coding harness with Ditto’s memory and deployment tools. Subnet 118 itself will incentivize miners to improve the underlying memory and agent system through DittoBench, measuring recall accuracy, response quality, tool use and speed, while the wider product increasingly consumes other Bittensor services such as Hippius for storage and Desearch for AI search. On the business side, Ditto currently charges primarily for inference while allowing users to bring their own API keys, with the longer-term plan focused on enterprise deployments, private company knowledge graphs, app-platform fees and eventually consumer hardware, with future enterprise revenue intended to fund alpha-token purchases and burns.
Big shoutout to Gordon Frayne for his incredible work on his TAO Pill podcast! His deep dives into Bittensor Subnets break down the decentralized AI landscape like few others can. Head over to his YouTube channel to catch the full series for top-tier analysis and insights.
Recorded in May 2026. Gordon Frayne speaks with Payton, CEO of Ditto, and Sebie about bringing Ditto, Bittensor Subnet 118, into the ecosystem and turning the product into what they describe as an agentic operating system for both individuals and teams. Payton explains that Ditto has already been developed for several years as a consumer-focused AI workspace, but the team had reached limits around engineering resources, economics and distribution, making Bittensor a natural fit for accelerating the underlying technology through decentralized competition. The core idea is to create a single contextual hub that sits across users’ chats, files, agents and workflows, allowing tools such as Claude, Codex, OpenClaw and Hermes to share the same persistent memory rather than forcing users to continually copy information between separate applications. Miners are being incentivised to improve Ditto’s underlying model harness, with an initial emphasis on memory, context management, speed, cost and eventually more specialized capabilities such as coding and agent orchestration, while increasingly difficult benchmarks are used to continually raise the performance baseline. The conversation explores how a stronger harness could allow open-source models such as DeepSeek to perform far beyond their default capabilities, potentially giving enterprises Claude-level usefulness at significantly lower cost while also allowing the entire system to run privately or even offline on their own infrastructure. They also discuss Ditto’s broader plan to integrate other Bittensor subnets for capabilities such as storage, compute, computer vision, lead generation and specialist agents, making Ditto a front-end orchestration layer through which users can access intelligence from across the ecosystem. Looking ahead, the roadmap includes shared organisations, projects, team workspaces, an orchestrator agent capable of creating and managing specialist agents, background tasks, enterprise licensing and increasingly sophisticated shared memory, while the commercial model is expected to remain low-cost with revenue used for Subnet 118 alpha buybacks. The longer-term vision is for Ditto to become a unified operating system where people and AI agents work together through a common memory and context layer, replacing the fragmented mix of chat apps, file systems and AI tools that users currently rely on.
Live in Ditto 1,342+ app integrations.
Browse by category or search directly. Connect Google Workspace, Notion, Linear, HackMD, whatever's in your stack. Multiple accounts supported.
Same memory layer underneath, so your agents stay coordinated across every tool.
What gets measured improves.
We know we've got work to do, and tracking our metrics shows us where to focus. Right now that's preparing for Ditto V2 launch which includes:
- new features
- redesigned app
- greatly improved user onboarding
Expect more updates like this.
Your history doesn't have to start in Ditto. Bring it from ChatGPT, Claude, Gemini or Perplexity, and Ditto reads the export straight from the ZIP.
Notes too. An Obsidian vault, a Joplin .jex, a folder of .md and .txt files, Google Keep out of Takeout, Apple Notes as Markdown...
Live in Ditto 1,342+ app integrations.
Browse by category or search directly. Connect Google Workspace, Notion, Linear, HackMD, whatever's in your stack. Multiple accounts supported.
Same memory layer underneath, so your agents stay coordinated across every tool.
What gets measured improves.
We know we've got work to do, and tracking our metrics shows us where to focus. Right now that's preparing for Ditto V2 launch which includes:
- new features
- redesigned app
- greatly improved user onboarding
Expect more updates like this.
"The future of AI agents isn’t giving every agent unrestricted access.
It’s managed tools: the router is the harness, the API key is the agent, your memory is the model, and sensitive actions go through an approval queue.
Safe delegation without giving up control." -
