ctxloom

Stop drawing the graph. Build agents as reactions to versioned, provable artifacts.

Most agent frameworks make you draw the graph: connect nodes, wire memory, declare control flow. But a knowledge question — "why did infra costs jump in Q2?" — needs Confluence + GitLab + CSV + calculations + verification, and the next question needs a different path. There is no universal graph to draw.

ctxloom flips the model. You describe what artifacts exist and what agents can do with them; the runtime derives what runs next from state changes. Agents react to events — there is no graph, no node pipeline.

pip install ctxloom

Runs offline, no API key needed — paste this straight into a .py file. Two agents, no graph edge declared between them — the second reacts because the first one's output exists, and the answer carries proof of where it came from:

from pydantic import BaseModel

from ctxloom import Budget, Consume, Context, Runtime, RuntimeResources, create_agent, produce


class Question(BaseModel):
    text: str


class Evidence(BaseModel):
    text: str


class Answer(BaseModel):
    text: str


DOCS = {
    "refund": "Refunds are available within 14 days of purchase.",
    "pricing": "The Pro plan is $49/month, billed annually.",
}


@produce(Evidence)
async def find_evidence(context, inputs, event, effects):
    question = next((a for a in inputs if isinstance(a.data, Question)), None)
    if question is None:
        return None
    hit = next((v for k, v in DOCS.items() if k in question.data.text.lower()), None)
    if hit is not None:
        effects.create(Evidence(text=hit))


@produce(Answer)
async def answer_from_evidence(context, inputs, event, effects):
    evidence = next((a for a in inputs if isinstance(a.data, Evidence)), None)
    if evidence is None:
        return None
    effects.create(Answer(text=evidence.data.text)).link("supported_by", evidence)


search_agent = create_agent("search", consumes=[Consume(Question)], produces=[find_evidence])
answer_agent = create_agent("answer", consumes=[Consume(Evidence)], produces=[answer_from_evidence])

ctx = Context(resources=RuntimeResources())
runtime = Runtime(ctx, agents=[search_agent, answer_agent], budget=Budget(max_runs=10))

ctx.create(Question(text="what's your refund policy?"))
runtime.run()  # search_agent and answer_agent both react — nobody wired them together

answer = ctx.latest(Answer)
evidence = ctx.related(answer.id, "supported_by")[0]
print(answer.data.text)                     # "Refunds are available within 14 days of purchase."
print("supported_by:", evidence.data.text)  # provenance you can trace, not just a string in a log

How it works

ARTIFACT CREATED / UPDATED
       │
       ▼
     AGENTS REACT ──self.effects──► Effects ──compile──► Patch
       ▲                                                      │
       └──────────────────────────────────────────────────────┘
                                                        Context v+1

A produce writes what should change (self.effects.create/update/link/ask) and returns None; the runtime compiles the effect set into one atomic Patch and moves the context to the next version. The event that wakes an agent is derived from that same change — the causal chain can never drift from the actual state.

What makes it different

Reactive. Deterministic. Accountable.

Full breakdown, including where ctxloom is not the right choice: docs/en/comparison.md.

Core primitives

• Context — versioned working state, git-like commits, diff/rollback/merge.
• Artifact — a first-class typed object (Claim, Evidence, Answer), not a string blob.
• Effects — an agent states its change via self.effects.create/update/link/ask; the runtime compiles it.
• Patch — the compiled, validated change-set applied as one atomic commit.
• Agent — a thin container declaring consumes/produces; logic lives in a Produce.
• Source — retrieval is a capability: vector search is one strategy; direct API, keyword, SQL, filesystem are equally first-class.
• Provenance — every derived artifact links to what produced it (Answer —supported_by→ Claim —derived_from→ Evidence —extracted_from→ Doc).
• HITL — humans as effects.ask(...) → PendingQuestion, answered via effects.resume(...) like any agent.

In the box

• Deterministic by design — calculations over structured data, honest None fallbacks instead of hallucinated answers; the model reasons, never "knows".
• Observability — every run traces agent spans, reads/writes, LLM calls, tokens: SQLite store + web dashboard, exportable to Langfuse/Postgres (async sinks).
• Budgets & replanning — cap by runs/time/iterations/tool-calls, replan on decline.
• Branching & replay — context.branch(), three-way merge(), deterministic ReplayLLM, all for audit and safe alternative states.
• Sessions — SessionStore over FileKVBackend/SQLiteKVBackend (and PostgreSQLKVBackend) for durable chat memory.
• Web layer — ChatAssistant + create_chat_router mount a canonical SSE chat on your FastAPI app; errors degrade to a logged fallback, never a 500.
• Recipes — find/find_all (typed lookup in inputs), fan_out_sources, materialize_doc, StatusMachine, WindowSummarizer/WindowPruner (bounded conversation memory), change→rebuild rollback helpers, Skill/match_skills (Claude-Skills-shaped instructions, keyword-triggered) — pure and LLM-free, except the summarizer, which takes your callback.
• Viz & CLI — Mermaid blueprint/context_to_mermaid/trace_to_mermaid; ctxloom with graph/context/trace/replay/branch.

Run a demo

Offline-capable, no API keys required (deterministic fallbacks):

uv run python ./examples/llm_ladder/level1.py  # the simplest LLM turn (offline too)
uv run python ./examples/repair/web.py         # room renovation: plan, estimate, CSV export
uv run python ./examples/devops/web.py         # HITL ops assistant + trace dashboard

Classic-pattern ports run as one-liners too: python -m examples.{reflection,map_reduce,supervisor,summarize,time_travel,adaptive}.main.

Examples (in-repo, not shipped)

• knowledge — multi-source chat: search → evidence → claim verification → answer, with CSV calculation.
• research — goes to the web (WebSource): lazy page fetch → evidence → verified claims → answer with URL provenance.
• medic-lab — hypothesis laboratory: competing hypotheses scored, HITL steering, honest report.
• devops — HITL tool agents + LLM tool router + trace dashboard.
• repair — budget-aware replanning (chat/data in Russian by design).
• forklab — deterministic branch & merge: two strategies on their own forks, three-way merge.
• llm_ladder — the workflow from one LLM call to state-changing patches (3 levels).
• adaptive — hybrid scheduler: rule filters + deterministic rank + LLM tie-break + rank_limit.
• {reflection,map_reduce,supervisor,summarize,time_travel} — canonical ports (see port-matrix).

Documentation

• English · Русский — concepts, sources, providers, recipes, patterns, observability, eval, branching, replay, viz/CLI, API.
• Why ctxloom — the design argument: why effects, why no graph, why determinism.
• Comparison — ctxloom vs LangGraph/CrewAI, feature by feature, and when not to use ctxloom.
• Tutorial · llm-ladder — learn the workflow.
• docs/constitution.md — the full design rationale and invariants.

Development

uv sync --extra dev --extra web
.venv/bin/python -m pytest
.venv/bin/mypy
.venv/bin/ruff check

License

MIT — see LICENSE.