simagents 0.2.1

Lightweight multi-agent orchestration with prompt-planning workflows

simagents

simagents is a lightweight Python framework for building multi-agent workflows with:

• 🔁 Linear, parallel, and loop orchestration modes
• 🤖 Agent-level model configuration
• 🧭 Prompt-planning friendly task chaining
• 📝 Safe decision logs (reasoning summaries)
• 💾 Retry/backoff + run artifact persistence
• ⚡ Optional exact-response caching to reduce repeated LLM calls

Why simagents (vs broader frameworks)

• Workflow-first 🔁: orchestration mode is a first-class setting (linear, parallel, loop)
• Prompt-planning native 🧭: easy to build research → prompt-plan → execution flows
• Simple API ✨: define agents + tasks + workflow, then run
• Production-lite defaults 🧰: retries, logs, artifact folders, decision logs, optional cache

Install (local)

From the simagents/ folder:

pip install -e .

For tests/dev:

pip install -e ".[dev]"

Environment variables

simagents supports multiple provider adapters via the OpenAI SDK-compatible interface:

• OpenAIProvider
• OllamaProvider
• OllamaCloudProvider
• GroqProvider
• TogetherProvider
• OpenRouterProvider
• AnthropicProvider (Claude)
• OpenAICompatibleProvider (custom base URL)

Base env vars:

SIMAGENTS_API_KEY=your_key
SIMAGENTS_BASE_URL=https://api.openai.com/v1

If you do not pass a provider explicitly to EasyOrchestrator, simagents uses OpenAICompatibleProvider() by default. In that case, set either:

export SIMAGENTS_API_KEY="your_api_key_here"
export SIMAGENTS_BASE_URL="https://api.openai.com/v1"

or set OpenAI's standard key:

export OPENAI_API_KEY="sk-..."

For a one-off command without saving the variables in your shell session:

OPENAI_API_KEY="sk-..." python examples/research_prompt_plan.py

For an OpenAI-compatible provider such as Ollama Cloud:

SIMAGENTS_API_KEY="your_ollama_key" \
SIMAGENTS_BASE_URL="https://ollama.com/v1" \
python examples/research_prompt_plan.py

Note: the model names in your agents must match the provider you use. The bundled examples/research_prompt_plan.py uses gpt-4o-mini, which is an OpenAI model. If you run it against Ollama Cloud, change the example agent models to an Ollama Cloud model such as gpt-oss:120b-cloud.

Fallback key env var:

• OPENAI_API_KEY

Provider-specific common keys:

• OLLAMA_API_KEY, OLLAMA_BASE_URL
• OLLAMA_CLOUD_API_KEY, OLLAMA_CLOUD_BASE_URL (defaults to https://ollama.com/v1)
• GROQ_API_KEY
• TOGETHER_API_KEY
• OPENROUTER_API_KEY
• ANTHROPIC_API_KEY

Web search provider keys:

• TAVILY_API_KEY
• GOOGLE_API_KEY
• GOOGLE_CSE_ID

Claude model examples:

• claude-4-6-sonnet-latest
• claude-4-7-opus-latest

Quickstart

from simagents import AgentSpec, EasyOrchestrator, RunConfig, TaskSpec, WorkflowSpec
from simagents.core.models import WorkflowMode
from simagents.llm import AnthropicProvider, OpenAIProvider

agents = [
    AgentSpec(name="researcher", role="Research specialist", model="gpt-4o-mini"),
    AgentSpec(name="writer", role="Technical writer", model="gpt-4o-mini"),
]

tasks = [
    TaskSpec(name="research", agent_name="researcher", prompt_template="Research: {input}"),
    TaskSpec(name="final", agent_name="writer", prompt_template="Write post using: {research}"),
]

workflow = WorkflowSpec(mode=WorkflowMode.LINEAR)
run_config = RunConfig(output_dir="runs", save_artifacts=True)

orch = EasyOrchestrator(
    agents=agents,
    tasks=tasks,
    workflow=workflow,
    run_config=run_config,
    provider=OpenAIProvider(),
)
result = orch.run(input_text="How AI is changing bioinformatics")
print(result.final_output)
print(result.decision_log)

# Claude usage (swap provider)
# orch = EasyOrchestrator(
#     agents=agents,
#     tasks=tasks,
#     workflow=workflow,
#     run_config=run_config,
#     provider=AnthropicProvider(),
# )

LLM response caching

simagents can cache exact LLM invocations to reduce token/API usage when agents repeat the same work.

Caching is disabled by default so prompt iteration remains fresh and unsurprising. Enable it in RunConfig:

run_config = RunConfig(
    output_dir="runs",
    save_artifacts=True,
    cache_enabled=True,
    cache_dir=".simagents_cache",
    cache_ttl_seconds=None,  # optional; set seconds to expire old entries
)

Cache keys include:

• provider class name
• provider base URL, when available
• model
• temperature
• full rendered prompt
• internal cache version

This means caching is safe and deterministic for exact repeats. If the prompt, model, temperature, or provider changes, simagents treats it as a new call.

When caching is enabled, the decision log notes whether a task stored fresh output or reused cached output.

Orchestration modes

• WorkflowMode.LINEAR: run tasks one by one
• WorkflowMode.PARALLEL: run tasks concurrently
• WorkflowMode.LOOP: rerun full task chain until stop keyword appears or max iterations reached

In PARALLEL mode, TaskSpec.depends_on is respected as a dependency graph.

Loop controls:

• WorkflowSpec.max_iterations
• WorkflowSpec.stop_condition_keyword

Flagship example: research + prompt planning

From the simagents/ folder, run:

python examples/research_prompt_plan.py

If you installed simagents from the workspace root and want to run the example directly from this repository, you can also run:

python simagents/examples/research_prompt_plan.py

Running the example with API keys

OpenAI, using the example as-is (gpt-4o-mini):

OPENAI_API_KEY="sk-..." python simagents/examples/research_prompt_plan.py

OpenAI-compatible endpoint:

SIMAGENTS_API_KEY="your_key" \
SIMAGENTS_BASE_URL="https://api.openai.com/v1" \
python simagents/examples/research_prompt_plan.py

Ollama Cloud one-liner:

SIMAGENTS_API_KEY="your_ollama_key" \
SIMAGENTS_BASE_URL="https://ollama.com/v1" \
python simagents/examples/research_prompt_plan.py

For Ollama Cloud, update the example models first, for example:

model="gpt-oss:120b-cloud"

You can also export keys once per terminal session:

export OPENAI_API_KEY="sk-..."
python simagents/examples/research_prompt_plan.py

This example demonstrates:

1. Research agent gathers structured topic context
2. Planner agent turns research into a high-quality prompt blueprint
3. Writer agent executes using that prompt plan

Output artifacts

When save_artifacts=True, each run creates:

• runs/run-<timestamp>/decision_log.md
• runs/run-<timestamp>/final_output.md
• one markdown file per task name

Web search providers (Tavily, DuckDuckGo, Google CSE)

You can use pluggable search providers:

from simagents import (
    TavilySearchProvider,
    DuckDuckGoSearchProvider,
    GoogleCustomSearchProvider,
    format_search_results,
)

# Tavily
tavily = TavilySearchProvider()  # needs TAVILY_API_KEY
print(format_search_results(tavily.search("AI bioinformatics", max_results=3)))

# DuckDuckGo (instant answer + related topics)
ddg = DuckDuckGoSearchProvider()
print(format_search_results(ddg.search("AI bioinformatics", max_results=3)))

# Google Custom Search JSON API
google = GoogleCustomSearchProvider()  # needs GOOGLE_API_KEY + GOOGLE_CSE_ID
print(format_search_results(google.search("AI bioinformatics", max_results=3)))

Lifecycle hooks

You can attach optional hooks for observability/instrumentation:

• on_step_start(step_name)
• on_step_end(step_name, output)
• on_error(step_name, exception)

Testing

pytest -q

License

MIT License. See LICENSE for details.

---

Built with <3 on a cloudy Sunday.