HypSwarm

Many agents. One record.
Run one agent or a swarm of them on a research, engineering, trading, knowledge or optimisation task, with a HypGraph record as the only shared state.
hypswarm.com · for agents · for engineers · examples · the design · decisions · HypGraph

Alpha.

The record schedules the work (orders from its question board), accepts it (predicates the record checks itself), stops what became pointless (premises that fell), re-runs what authors computed (the instrument), and writes the report. Models are called only to do work and to judge exceptions. A solo scientist and an orchestrated swarm are the same program with different seats — and a trading desk, a software team or a research assistant are the same program with a different profile.

import asyncio
from hypswarm import Task, Budget, run
from hypswarm.providers import OpenAICompatible

task = Task(title="Does A beat B",
            mission="Two models were scored on the same items (items.csv). Find out whether A beats B.",
            questions=[("Q1", "What is A's accuracy?"), ("Q2", "Is A significantly better than B?")],
            files={"items.csv": open("items.csv", "rb").read()},
            profile="scientific")                       # or software · trading · knowledge · optimization
model = OpenAICompatible("deepseek-ai/DeepSeek-V4.1-Flash", base_url="https://api.deepinfra.com/v1/openai")
result = asyncio.run(run(task, None, model, record="study.hypgraph", budget=Budget(max_calls=150, max_usd=1.0)))
print(result.report)        # the facts from the record, the lead's interpretation under them, the answers

seats=None takes the profile's roster. Or from the shell: hypswarm run --task task.toml --profile software --record study.hypgraph.

Profiles: what counts, per discipline

A profile is a HypGraph governance profile — what the record will accept and refuse, and the discipline's own workflow: the stages a question passes through and the gate out of each — plus the seats and tools that usually work under it. It is pinned on the investigation and versioned, so a record written under trading@2 means the same thing in a year. A name (trading) is its newest version; trading@1 pins the pre-workflow one.

The stage is derived, never set. The record reads a question's stage off what has been written (hypgraph's stage()); the scheduler issues that stage's order — the gates up to the next review as its done-when, to the seat the stage names — and a review stage goes to a reviewer on the node the previous gate produced. A question is complete when the record says so. An honest "it can't be answered" — a checked claim of kind inconclusive, derived from what the search did find — completes a question at any stage; declining a thesis is a decision (kind=no_position).

Every profile adds a paragraph of its rules to each seat's preamble and tells the scheduler what "complete" means (completion()); hypswarm profiles lists them. Adding a discipline is a Profile with a Workflow in HypGraph plus a roster here — the scheduler does not change. The method each workflow is taken from is quoted in HypGraph's docs/workflows/.

Workflows, per profile

What each stage's order asks a seat to write, and what the record refuses on the way. Italic steps are mechanical — no model is called.

scientific — question → a hypothesis with predictions → an experiment that tests it with the expected outcome, before any run → register_run executing it (run + result + claim, numbers bound) → instrument re-runs the code → review by the critic (judges; recomputes only a contradiction) → question settled → report from the facts, linted. A corrected data file puts what rested on it at risk; a check on stale evidence doesn't count.

software — requirement → the change as an artifact kind=diff (base → head sha, files) → register_run with the exact test command as code and uses the diff → instrument runs the tests on that head (tests_pass_current) → claim "passes", accepted only now → review by someone who wrote neither the claim nor the diff → requirement verified. A newer head supersedes the old diff: the tests go stale and the gate reopens.

trading — market question → the thesis as a hypothesis → the backtest preregistered as an experiment (windows, costs, configurations to try) → register_run with oos_return, fees, slippage, trials → instrument re-runs the backtest → review by the risk seat → a decision based_on the claim: kind=position with size, stop, review_date, paper — or kind=no_position → monitored: past its review date with nothing newer, the position is flagged.

knowledge — question → (the search as an experiment, optional) → source node (citation, URL or DOI, the passage, retrieved, kind) part_of the question → claim that cites it → review by the fact-checker, who opens the source → all checked. A withdrawn source puts every claim citing it at risk. No source? The search's own result ("read 3 of 3 pages; 0 mention it") and a claim kind=inconclusive derived from it settle the question honestly.

optimization — question + the instance as an artifact kind=instance → the model as a hypothesis → the solve configuration as an experiment → register_run that uses the instance, with objective, feasible, bound or gap → instrument re-runs the solve → review of the result by the verifier against the instance → claim "best found, gap x", or "optimal" only with a bound → review of the claim → incumbent verified. A changed instance supersedes the solve.

Each of these is one of the examples, runnable for free; the trace shows every stage order and the refusals on the way.

Watching a run

Every run that writes a record file also writes <record>.telemetry.jsonl: each model call, tool call, work order and scheduler decision as a span, with tokens, cost and timing. Serve the record and open the HypGraph UI's Live view to watch the swarm work: which seat is calling the model or running a tool right now, what each has spent, and a trace of where a work order's time went.

hypgraph --db desk.hypgraph serve        # then run the swarm against desk.hypgraph and open Live
hypgraph --db desk.hypgraph export --telemetry -o desk.bundle.json   # a bundle that replays the run with no server

See it on five finished runs, one per profile: hypgraph.com/examples.

otel=True sends the same spans and metrics to OpenTelemetry (pip install 'hypswarm[otel]'). Prompts, replies and tool arguments stay out unless capture_content=True.

When the world changes

run(..., watch=True) keeps the record in step with the files and pages the run used, with no model call: a file whose bytes changed supersedes its artifact, a commit is a new head of the working tree, and (with refetch=) a page that no longer says what a source quotes makes the source stale. Results computed from the old state go STALE; the instrument re-runs them on the current files; the ones that still reproduce come back on their own, and only the answers whose evidence really moved are redone. Nobody has to tell the team that a teammate pushed a commit or the vendor corrected a file — the record holds the hashes.

When the lead never gets a report past the record's check, the harness writes one from the record itself — the facts and the answers the current claims give, labelled as assembled (res.report_source == "assembled").

res = await run(task, None, model, record="desk.hypgraph", folder="work/", reproducer=my_sandbox,
                watch=True, refetch=my_fetch)

What you get

• A scheduler that makes no model calls. It issues orders to idle seats from the question board, settles every returned order whose predicates are met, reworks the rest twice, cancels what a fallen premise made moot, and brings in the lead seat only at exceptions and for the report.
• Workers that hold no state. A worker's context is its order and its own turns; every reply from the record tells it whether its order is done. It can be killed: the lease expires and another takes the order, and a run that died can be resumed from the record (run(..., resume=inv)).
• The instrument reproduces; reviewers judge. register_run(code=…) keeps the code; before any review the host's sandbox re-runs it and the record files a lab run — values confirmed, not printed, or contradicted. A reviewer recomputes only a contradiction.
• Claims that can't be flattered. Every claim is checked as it is written (HypGraph's refusal rules and the profile's), the record binds the numbers it already holds, one review settles a question, and the final report is linted against the record before it is accepted.
• A budget that is a rule. Calls and dollars, per run and per order, counted including calls in flight, and shown to every seat.
• Escalation the record decides. A seat may name a deeper model (escalate_to); it is used from the first time the record sends that seat's work back NOT MET — never on the model's own say-so.
• Files the record vouches for. file_tools reads a file only when its bytes match a live artifact; a withdrawn file is refused with the record's reason.
• An audit trail you didn't write. Orders are task nodes, acceptance is a decision, the run files its own provenance: the record is the log, and it verifies by hash.

Seats

Three roles, because they are what the scheduler can check. Jobs are seats: a name, the words it takes, the toolkits it may use, a model and optionally a deeper one to escalate to.

Tools

Tools are named bundles from a registry. hypswarm ships files; anything that executes code or reaches a network is the host's to register, so the swarm never runs what the host did not choose — and a profile that asks for a bundle the host hasn't registered fails loudly rather than running without it.

from hypswarm import toolkits
toolkits.register("python", lambda ctx: [Tool.from_function(my_sandboxed_python)])
result = await run(task, None, model, record="r.hypgraph", folder="work/", toolkits=["files", "python"])

People and other agents as seats

A seat can be external: no model behind it, just a name in the record. The scheduler issues it orders like any other seat and waits; whoever holds that name works the order through the HypGraph MCP server and returns it:

[[seats]]
name = "nicolas"
role = "worker"
external = true

hypgraph --db study.hypgraph --as nicolas mcp     # in Claude Code, Cursor, or any MCP client
# my_work → the order; register_run / add / review → the writes; return_work when the reply says every predicate is met

The record can't tell a person from an agent except by name, and it holds them to the same rules.

Bring your own agent

A seat can be held, in-process, by a whole agent from another framework — your prompts, your tools, your loop — and still be governed by the record: same orders, same refusals, same instrument re-run, same budget.

from hypswarm.bridges.langchain import langchain_agent          # LangChain / LangGraph: any Runnable over {"messages"}
from hypswarm.bridges.pydantic_ai import pydantic_ai_agent       # PydanticAI Agent
from hypswarm.bridges.crewai import crewai_agent                 # CrewAI Crew

seats = [Seat("analyst", "worker", agent=langchain_agent(lambda tools: create_react_agent(llm, tools))),
         Seat("stats", "worker", agent=pydantic_ai_agent(lambda tools: Agent("openai:gpt-4.1-mini", tools=tools))),
         Seat("critic", "reviewer", agent=crewai_agent(lambda tools, order: Crew(...))),
         Seat("lead", "lead")]                                   # a model, e.g. Anthropic(...) or Bedrock(...)

A bridge hands the agent the order, the seat's system prompt and the record's tools in the framework's own tool shape, caps the framework's loop at the calls left in the run, and reports what it used. If the agent doesn't call return_work itself, HypSwarm returns the order with its final answer as the note. Your own framework is one async def agent(order: AgentOrder) -> AgentOutcome away (ADR-009; examples/10-custom-seat-and-toolkit). examples/06-mixed-swarm runs LangChain, PydanticAI, CrewAI, Claude / Bedrock and a person on one record.

One record server, many workers

Workers don't have to share a process. Serve the record, bind writes to tokens, and point the run at it:

hypgraph --db study.hypgraph serve --writable --no-browser --token "$SWARM_TOKEN=harness,quant,risk,lead" --token "$NIC_TOKEN=nicolas"
hypswarm run --task task.toml --profile trading --record-url http://127.0.0.1:8765

In Python, run(..., record=HttpRecord(url, token=...)). A token is bound to the authors it may write as; the harness name is reserved, so an outside seat can hold its own name and nothing else. Every write still goes through the record's checks, the file still verifies by hash, and workers on other processes or machines are stateless.

Providers

The core ships OpenAICompatible (DeepInfra, OpenAI, NVIDIA NIM, vLLM, Ollama, most gateways) and a scripted StubModel for tests. Extras:

Any object with async complete(messages, tools) -> Reply is a model.

Examples

Ten runnable examples in examples/, each free on scripted models and cents on a real provider: one per profile (an eval report, tests the instrument runs, a paper trading desk, a sourced brief, a solve with a bound), the mixed swarm (LangChain + PydanticAI + CrewAI + Claude / Bedrock + a person), a person as reviewer over MCP, a served record with tokens, stop-and-resume with a file withdrawn mid-run, and a custom seat and toolkit from scratch. Several make a deliberately wrong first write so the trace shows the refusal and the honest rewrite.

Installing, today

HypSwarm needs HypGraph's coordination layer and its workflows, and the claim checks that accept an honest sentence on the first try, which ship with hypgraph 0.7.2:

pip install --pre hypswarm                 # pulls hypgraph>=0.7.2; --pre while it is an alpha

Status

Alpha. Built from the harness the HypGraph harmony study runs agent teams on, and measured there as a preregistered condition before anything is claimed for it. What's built: the scheduler, workers, five profiles with rosters, the tool registry, external seats over MCP, the record server with token-bound writes, resume, escalation, instrument reproduction, verified file tools, the OpenAI-compatible, Anthropic, Bedrock and LangChain adapters, agent seats with bridges for LangChain / LangGraph, PydanticAI and CrewAI, ten runnable examples, and a test suite that runs every scheduler path — and every bridge — on scripted models in seconds. Every design decision is in docs/decisions.

Made by Nicolas Houghtaling at VoodooMods LLC. Apache-2.0.