How it works

Thronos

Agent operations platform. A team of AI agents that works on a schedule. They read a company’s systems, draft the work and bring every decision to a person to approve.

Live · privateVisit the site ↗

Fig.A team of AI agents that does a company’s routine work on a schedule, remembers what it learns, and brings every decision to a person.

Architecture

AWeb· 5Stage · Decisions · Library · Live socket · Slack
- StageThree.js roomUI
  UI · Three.js room
  The home screen: each agent a node with live presence, lines where one hands work to another, an artifact shelf and a chat dock.
  Talks to
  ← Live socket (live)
- Decisionsj · k · a · rUI
  UI · j · k · a · r
  One queue for every pending approval, gate and proposal, triaged from the keyboard, with a log of what was settled today.
  Talks to
  → Engine API (approve · reject)
- LibraryartifactsUI
  UI · artifacts
  Every typed artifact an agent has produced, addressable by key.
- Live socketsocket.ioREALTIME
  REALTIME · socket.io
  One shared connection per app; the Stage, chat and live thought streams all listen on it.
  Talks to
  ← Redis (pub/sub)
  → Stage (live)
- Slackfast-path chatEXTERNAL
  EXTERNAL · fast-path chat
  Talk to the agents from Slack, answered by the cheaper runtime.
  Talks to
  → Kernel (chat)
BEngine· 16Engine API · Signals · Heartbeat · Orchestrator · 14 sub-agents · Kernel · Team templates · Governance · Agent runner · Task loop · Event mesh · MCP registry · Autoresearch · Self-healing · Costs · Integration hub
- Engine APInode:httpSERVICE
  SERVICE · node:http
  The engine’s HTTP API, on Node’s own http module rather than a framework.
  Talks to
  ← Decisions (approve · reject)
  → Governance
- SignalsGitHub · Linear · healthJOB
  JOB · GitHub · Linear · health
  Gatherers that collect what changed (code, tickets, business numbers, tool health, memory) before each tick.
  Talks to
  → Heartbeat (what changed)
- HeartbeatBullMQ · 12 lanesJOB
  JOB · BullMQ · 12 lanes
  Health hourly, the inbox every two hours, reviews each morning, weekly analysis, nightly research. Lanes that share resources never overlap, and a circuit breaker pauses an agent that keeps failing.
  Talks to
  ← Team templates (seeds)
  ← Signals (what changed)
  → Orchestrator (tick)
  → Kernel (cheap lanes)
  → Autoresearch (nightly)
- OrchestratorOpus · Agent SDKAI
  AI · Opus · Agent SDK
  The top-level agent. It reads the signals, opens and claims a task, and delegates to the specialists.
  Talks to
  ← Heartbeat (tick)
  → 14 sub-agents (delegates)
  → Task loop (claims)
- 14 sub-agentsHaiku · Sonnet · OpusAI
  AI · Haiku · Sonnet · Opus
  Sentinel, Analyst, Scribe, Custodian, Oracle, Architect, Strategist and the rest, each with its own model, tools, budget and yes-or-no checks.
  Talks to
  ← Orchestrator (delegates)
  → Agent runner
  → Governance (proposals)
  ← Autoresearch (prompts)
  ← Self-healing (revisions)
  → Anthropic · OpenRouter
- KernelOpenRouter · cappedAI
  AI · OpenRouter · capped
  A second, cheaper runtime for the frequent lanes, with a hard budget per tick, a turn limit, and only the MCP servers that lane needs.
  Talks to
  ← Heartbeat (cheap lanes)
  → MCP registry (subset)
  ← Slack (chat)
- Team templatesfiveSERVICE
  SERVICE · five
  Seeds a team, its schedule and its autonomy defaults in one action.
  Talks to
  → Heartbeat (seeds)
- Governanceapprovals · autonomySERVICE
  SERVICE · approvals · autonomy
  Approval records and the autonomy mode (supervised, assisted or full) that decides what is approved automatically. New agents are capped at 20 per workspace and three a day.
  Talks to
  ← Engine API
  ← 14 sub-agents (proposals)
- Agent runnerone run, start to endSERVICE
  SERVICE · one run, start to end
  Checks the budget, loads the agent’s past experience, filters its tools, runs it with a live thought stream, retries on failure, then writes the run to memory.
  Talks to
  ← 14 sub-agents
  → MCP registry (tools)
  → Event mesh (output)
  → Mnemos (write · recall)
  → Costs (checks)
- Task loopclaim · settleSERVICE
  SERVICE · claim · settle
  The work record. The engine claims a task before delegating and settles whatever is still open when a run ends, so the record is right even if an agent forgets.
  Talks to
  ← Orchestrator (claims)
- Event meshagent → agentREALTIME
  REALTIME · agent → agent
  Each run’s output is classified and published, so a Sentinel alert lands in the Orchestrator’s triage and an Architect deploy sets off a Sentinel check.
  Talks to
  ← Agent runner (output)
  → Redis (publish)
- MCP registry~60 built-in toolsAI
  AI · ~60 built-in tools
  Assembles each agent’s tools from the healthy servers, minus the tools that agent is blocked from.
  Talks to
  ← Agent runner (tools)
  ← Kernel (subset)
  → Dev tools (MCP)
  → Company systems (MCP)
  ← Integration hub (compiled)
- Autoresearch03:00 nightlyJOB
  JOB · 03:00 nightly
  Simulates five outputs, grades them on yes-or-no checks, mutates the prompt, simulates five more, and promotes only at five percent better. Every experiment is logged.
  Talks to
  ← Heartbeat (nightly)
  → 14 sub-agents (prompts)
- Self-healingbinary evalsJOB
  JOB · binary evals
  Scores every run and drafts a prompt revision when quality drops. A revision may change at most a fifth of the prompt, and after three failed revisions in a row a person is asked.
  Talks to
  → 14 sub-agents (revisions)
- Coststick · day · monthSERVICE
  SERVICE · tick · day · month
  Token budgets per agent per tick, per day and per month, under a global cap.
  Talks to
  ← Agent runner (checks)
- Integration hubspec → MCPSERVICE
  SERVICE · spec → MCP
  Compiles a REST or GraphQL connector from its spec into an in-process MCP server at install time.
  Talks to
  → MCP registry (compiled)
  ← Any API (spec)
CMemory· 1Mnemos
- Mnemosmemory serviceSERVICE
  SERVICE · memory service
  The memory, as its own service: written after every run, consolidated on a schedule, searched before the next. Drawn in full below.
  Talks to
  ← Agent runner (write · recall)
  → Neo4j
  → PostgreSQL (vectors)
DData· 3PostgreSQL · Neo4j · Redis
- PostgreSQLPrisma · pgvectorSTORE
  STORE · Prisma · pgvector
  Agents, runs, tasks, approvals, budgets and experiments, plus the memory’s embeddings.
  Talks to
  ← Mnemos (vectors)
- Neo4jknowledge graphSTORE
  STORE · knowledge graph
  The graph of what the agents have learned.
  Talks to
  ← Mnemos
- RedisBullMQ · pub/subSTORE
  STORE · BullMQ · pub/sub
  The heartbeat’s queues and the event channels.
  Talks to
  ← Event mesh (publish)
  → Live socket (pub/sub)
EReach· 4Anthropic · OpenRouter · Dev tools · Company systems · Any API
- Anthropic · OpenRoutermodelsEXTERNAL
  EXTERNAL · models
  Claude through the Agent SDK; other models through OpenRouter.
  Talks to
  ← 14 sub-agents
- Dev toolsin-process MCPEXTERNAL
  EXTERNAL · in-process MCP
  GitHub, Linear and Railway: code, tickets and deploys.
  Talks to
  ← MCP registry (MCP)
- Company systemsHTTP MCPEXTERNAL
  EXTERNAL · HTTP MCP
  C0, ContentSphere and OutreachSphere, read through their own MCP servers.
  Talks to
  ← MCP registry (MCP)
- Any APIREST · GraphQLEXTERNAL
  EXTERNAL · REST · GraphQL
  Any API added through the hub.
  Talks to
  → Integration hub (spec)

Connections · 29

→ request · ⇢ webhook, queue or schedule

Decisions → Engine API · approve · reject
Engine API → Governance
Team templates → Heartbeat · seeds
Signals → Heartbeat · what changed
Heartbeat ⇢ Orchestrator · tick
Heartbeat ⇢ Kernel · cheap lanes
Heartbeat ⇢ Autoresearch · nightly
Orchestrator → 14 sub-agents · delegates
Orchestrator → Task loop · claims
14 sub-agents → Agent runner
Agent runner → MCP registry · tools
Agent runner ⇢ Event mesh · output
Agent runner → Mnemos · write · recall
Agent runner → Costs · checks
Kernel → MCP registry · subset
Slack → Kernel · chat
14 sub-agents → Governance · proposals
Autoresearch ⇢ 14 sub-agents · prompts
Self-healing ⇢ 14 sub-agents · revisions
MCP registry → Dev tools · MCP
MCP registry → Company systems · MCP
Integration hub → MCP registry · compiled
Any API → Integration hub · spec
Mnemos → Neo4j
Mnemos → PostgreSQL · vectors
Event mesh ⇢ Redis · publish
Redis ⇢ Live socket · pub/sub
Live socket ⇢ Stage · live
14 sub-agents → Anthropic · OpenRouter

Inside: the memory

Memory isn’t a log. Everything an agent produces is read for what it’s about, merged with what is already known, consolidated every six hours, faded when unused, and searched by meaning, keyword and connection when it’s needed again.

AIn· 5Run output · Experience · Documents · From tools · Scenarios
- Run outputsuccessful runsSERVICE
  SERVICE · successful runs
  What an agent produced, sent to memory after the run.
  Talks to
  → Extract
- Experienceevery runSERVICE
  SERVICE · every run
  A record of each run, failures included, so an agent can learn from what it has done.
  Talks to
  → Knowledge graph (Experience node)
- Documentsfiles · URLs · YouTubeSERVICE
  SERVICE · files · URLs · YouTube
  Outside material ingested on request.
  Talks to
  → Extract
- From toolsMCP ingestAI
  AI · MCP ingest
  Anything an agent chooses to remember through its memory tools.
  Talks to
  → Extract
- Scenariosworld modelsSERVICE
  SERVICE · world models
  A seed document becomes a model: actors with interests, tensions between them, dynamics and assumptions.
  Talks to
  → Knowledge graph (world model)
BWrite· 6Extract · Deduplicate · Score · Store · Embed · Connect
- Extractontology-awareAI
  AI · ontology-aware
  Pulls out entities and relationships using the known types and their properties. Runs on a local model when one is available, Haiku otherwise.
  Talks to
  ← Run output
  ← Documents
  ← From tools
  → Deduplicate
- Deduplicatespelling · meaningSERVICE
  SERVICE · spelling · meaning
  Matches each entity to one already known: by fuzzy name (bigram similarity of 0.75 or more), then by meaning.
  Talks to
  ← Extract
  → Score
- Scoreconfidence · half-lifeSERVICE
  SERVICE · confidence · half-life
  Confidence comes from the source; lifespan from the type: 7 days for a signal, 90 for a client, 365 for a principle. Use keeps a memory fresh.
  Talks to
  ← Deduplicate
  → Store
- Storemerge · never dropSERVICE
  SERVICE · merge · never drop
  Writes nodes idempotently with bounded concurrency. If extraction fails, the raw text is kept as an observation rather than lost.
  Talks to
  ← Score
  → Embed
  → Knowledge graph (MERGE)
- Embed1536 dimensionsAI
  AI · 1536 dimensions
  Embeds each node for search. In production a failed embedding is an error, never a placeholder that would quietly poison recall.
  Talks to
  ← Store
  → Connect
  → Embeddings
- Connectauto-linkSERVICE
  SERVICE · auto-link
  Links new nodes to related ones already in the graph.
  Talks to
  ← Embed
  → Knowledge graph (edges)
CStores· 2Knowledge graph · Embeddings
- Knowledge graphL0 raw · L1 distilledSTORE
  STORE · L0 raw · L1 distilled
  Nineteen core kinds of memory in Neo4j (new kinds can appear), from signals and bugs to decisions and principles, each fact able to say when it became true and when it stopped.
  Talks to
  ← Experience (Experience node)
  ← Scenarios (world model)
  ← Store (MERGE)
  ← Connect (edges)
  ← Compression (L0 → L1)
  ← Reflection (insights)
  ← Summaries
  ← Past experience (experiences)
- EmbeddingspgvectorSTORE
  STORE · pgvector
  The vectors behind semantic search, cleaned up when a node is archived so it can’t haunt results.
  Talks to
  ← Embed
  ← Decay (archive)
  ← Hybrid search (dense)
DKeep· 4Compression · Reflection · Summaries · Decay
- Compressionevery 6 hJOB
  JOB · every 6 h
  Clusters similar memories, checks the cluster really is similar (cosine 0.65 or more), and replaces it with one distilled L1 memory. High-value memories are never compressed.
  Talks to
  → Knowledge graph (L0 → L1)
- Reflectionafter compressionJOB
  JOB · after compression
  Reads 12 to 60 recent observations, decisions and experiences together and writes up to eight new insights or patterns.
  Talks to
  → Knowledge graph (insights)
- SummariesrollingJOB
  JOB · rolling
  Keeps a short summary on each node from its properties and its recent connections.
  Talks to
  → Knowledge graph
- DecaydailyJOB
  JOB · daily
  Recomputes each memory’s freshness from its half-life and last use; below 0.05 it is archived. A node with no timestamp is left alone, never archived on a guess.
  Talks to
  → Embeddings (archive)
ERecall· 6Past experience · HyDE · Hybrid search · Graph walk · Filters · Rerank
- Past experiencebefore each runSERVICE
  SERVICE · before each run
  The agent’s own past runs, ranked by overlap with the task at hand, added to its prompt so it builds a view from what it has actually done.
  Talks to
  → Knowledge graph (experiences)
- HyDEhypothetical answerAI
  AI · hypothetical answer
  For abstract questions, drafts a likely answer and searches with that instead. Direct lookups skip it.
  Talks to
  → Hybrid search
- Hybrid searchdense + keyword · RRFSERVICE
  SERVICE · dense + keyword · RRF
  Runs semantic and keyword search and fuses the two rankings.
  Talks to
  ← HyDE
  → Embeddings (dense)
  → Graph walk
- Graph walk2 hopsSERVICE
  SERVICE · 2 hops
  Expands each hit two steps through the graph, so related context comes along.
  Talks to
  ← Hybrid search
  → Filters
- Filterstime · type · truthSERVICE
  SERVICE · time · type · truth
  Drops archived and invalidated facts, and can answer as of a past date.
  Talks to
  ← Graph walk
  → Rerank
- Rerankagainst the questionAI
  AI · against the question
  A model reorders the shortlist against the original question, not the expanded one.
  Talks to
  ← Filters

Connections · 23

→ request · ⇢ webhook, queue or schedule

Run output → Extract
Documents → Extract
From tools → Extract
Experience → Knowledge graph · Experience node
Scenarios → Knowledge graph · world model
Extract → Deduplicate
Deduplicate → Score
Score → Store
Store → Embed
Embed → Connect
Store → Knowledge graph · MERGE
Embed → Embeddings
Connect → Knowledge graph · edges
Compression ⇢ Knowledge graph · L0 → L1
Reflection ⇢ Knowledge graph · insights
Summaries ⇢ Knowledge graph
Decay ⇢ Embeddings · archive
Past experience → Knowledge graph · experiences
HyDE → Hybrid search
Hybrid search → Embeddings · dense
Hybrid search → Graph walk
Graph walk → Filters
Filters → Rerank

Two flows

How an agent remembers

Walk through it and each step lights the parts it runs through in the drawings above.

By count

agent roles: 15
heartbeat lanes: 12
built-in tools: ~60
kinds of memory, each with its own lifespan: 19
team templates: 5
autonomy modes: 3

Notes

N.1Every kind of memory has its own lifespan. A signal fades in a week, a client in three months, a principle in a year, and anything that keeps getting used stays fresh.
N.2Memory can answer “what did we believe in March?” Facts carry when they became true and when they stopped, and recall can ask about a past date.
N.3If a model can’t make sense of something, it is still kept as a raw note. Nothing an agent learns is silently dropped.
N.4The agents react to each other. A Sentinel alert lands in the Orchestrator’s triage, and an Architect deploy sets off a Sentinel check.

Built by

I design, architect and build software products like this one.

naram@hey.com·See the rest of my work

More systems