barebear 0.2.1

The free agentic AI course. 12 lessons, 12 concepts, real models. Read the framework — it's the textbook.

The free agentic AI course.

12 lessons. 12 concepts. Real models. No magic. Read the framework — it's the textbook.

---

Three doors

For students — Try Lesson 1 in Colab. Your first real LLM call in 60 seconds. No card, no install.

For teachers — See the syllabus. A free, drop-in 12-week module on agentic AI for A-level CS, AP CS, IB CS, intro university CS, and bootcamps. The instructor guide covers classroom setup, pacing, and assessment.

For engineers — Read the framework. The whole thing is ~1,500 lines of Python. No magic, no DSLs, no metaprogramming. Open it.

---

Why barebear exists

In 2026, every CS department on earth is scrambling to teach AI, and the curriculum doesn't yet exist. Teachers are improvising slide decks the night before. Students are reading 50,000 lines of magic in popular agent frameworks, getting confused, and giving up — or worse, not getting confused, and shipping things they don't understand.

Most agent frameworks were written to demo well. BareBear was written to be read. It is small enough that a sixth-form student can read every line of it in one sitting, and end up understanding what an "agent" actually is.

It runs against real LLMs from lesson one — no fake models, no toy mocks. The default backend is OpenRouter (free tier, no card required). Everything you build with barebear, you build against the same LLMs you'd use in production.

The framework is also genuinely useful for shipping production agents — the same primitives that let a student understand the agent loop in twenty minutes let an engineer ship an auditable agent in an afternoon. But the audience this release is built for is the student and the teacher.

---

The 12-lesson syllabus

#	Concept	What you'll build
1	What an LLM is	A single chat-completion call. See tokens go in and out.
2	The agent loop	A one-tool agent. Watch perceive → think → act → observe.
3	Tool design	A multi-tool agent. Schemas, descriptions, when the model picks each.
4	State & memory	Short-term (messages) vs long-term (State dict). Persist across runs.
5	Budgets	Add a runaway-loop bug; watch the budget save you.
6	Policy & guardrails	Block a tool, classify side-effects, declare risk.
7	Checkpoints	An email-send agent that pauses for human approval.
8	Planning	bear.plan() before bear.run(). When each is the right move.
9	Reflection	Self-critique loop with Reflect — agent reviews its own work.
10	Multi-agent	A Bear that hires another Bear. Multi-agent in <40 lines.
11	Evaluation	Read a Report. Write tests for an agent.
12	Honest uncertainty	Make the agent flag what it does not know.

Lessons 1–12 ship as Jupyter notebooks. Lessons 7–12 also ship as plain Python scripts. Browse the full syllabus →

For teachers: classroom setup, per-lesson pacing notes, common student pitfalls, suggested assessments, and 3-/6-/12-week subset paths are all in the instructor guide.

---

Setup

OpenRouter (recommended — free, no install)

pip install barebear[openai]
export OPENROUTER_API_KEY="sk-or-..."   # grab a free key at https://openrouter.ai

Ollama (offline, local)

pip install barebear[openai]
# from https://ollama.com — install the binary, then:
ollama pull qwen2.5:3b
ollama serve

---

Quickstart for engineers

from barebear import Bear, Task, Tool, Policy, OpenRouterModel

def greet(name: str) -> str:
    return f"Hello, {name}!"

bear = Bear(
    model=OpenRouterModel(),
    tools=[Tool(name="greet", fn=greet, description="Greet someone by name")],
    policy=Policy(max_steps=5, max_cost_usd=0.05),
)

result = bear.run(Task(goal="Greet a user named Alice"), trace=True)
print(result.summary())

trace=True streams every loop turn to stdout — for a live narration of the agent's thinking. Drop it in production.

─── turn 1 ───
> calling model with 2 messages, 1 tools available
< tool call: greet(name="Alice")
→ greet returned: Hello, Alice!
─── turn 2 ───
< final answer: Hello, Alice! How can I help you today?

============================================================
status: completed  |  turns: 2  |  tokens: 187  |  cost: $0.0001
============================================================

Full quickstart with policy, custom system prompts, reflection, and multi-agent: docs/quickstart.md.

---

The 7 primitives

Primitive	What it does
Bear	The agent. Holds model, tools, policy, state. Runs tasks.
Task	A unit of work: a goal string, input dict, optional context, optional system prompt.
State	Explicit key-value store with snapshot history and change tracking.
Tool	A callable with declared risk, side effects, and approval requirements.
Policy	Constraints: step limits, cost caps, blocked tools, approval lists.
Checkpoint	A saved pause-point for human approval before high-risk actions.
Report	The run receipt: every step, token count, cost, assumptions, uncertainties.

Plus two helpers for advanced lessons:

Helper	What it does
Reflect	Three-method self-critique: critique, revise, run.
bear.as_tool(...)	Wraps a Bear so another Bear can call it. Multi-agent in one line.

That's it. No chains, no graphs, no planners, no routers. Just the bones.

---

Features

Policy-first tool execution

Every tool call passes through policy before it runs. Block tools, require approval, or reject external side effects — all in one declaration.

policy = Policy(
    max_steps=10,
    blocked_tools=["delete_account"],
    require_approval_for=["send_email"],
    allow_external_side_effects=False,
)

Budget tracking

Step counts, tool calls, token usage, and dollar cost — all tracked against limits you set. If a run hits its budget, it stops cleanly with a budget_exceeded status.

policy = Policy(max_steps=8, max_cost_usd=0.10, max_tokens=50000)

Checkpoint / approval gates

High-risk tools pause the run and create a Checkpoint. You approve or reject, then resume.

bear = Bear(
    model=model,
    tools=[
        Tool("send_email", fn=send_email, risk="high",
             side_effects="external", requires_approval=True),
    ],
    policy=Policy(require_approval_for=["send_email"]),
)

result = bear.run(task)

if result.status == "paused":
    checkpoint = bear.checkpoints.get(result.checkpoint_id)
    result = bear.resume(checkpoint, approved=True)

Side-effect staging

Tools declare side_effects="none", "internal", or "external". Policy can block external side effects entirely, so your agent can propose actions without executing them.

Tool("propose_patch", fn=propose, description="Suggest a code change",
     side_effects="none")
Tool("apply_patch", fn=apply, description="Apply the change",
     side_effects="external")

Run receipts

Every bear.run() returns a Report — a full trace you can print, serialise to JSON, or store for auditing.

result = bear.run(task)

print(result.summary())       # human-readable receipt
print(result.to_json())       # full JSON trace
print(result.total_cost_usd)  # what it cost
print(result.steps)           # list of every step taken

Honest uncertainty

Bears track what they don't know. Assumptions and missing information are first-class data in the report, not buried in log noise.

result = bear.run(task)
print(result.assumptions)     # ["Customer tier assumed from email domain"]
print(result.uncertainties)   # ["Could not verify account status"]

Multi-agent in one line

researcher = Bear(model=OpenRouterModel(), tools=[search_tool])
writer = Bear(
    model=OpenRouterModel(),
    tools=[researcher.as_tool(name="research", description="Research a topic")],
)
writer.run(Task(goal="Write a brief on UK A-level CS reform."))

Reflection in three calls

from barebear import Reflect, OpenRouterModel

reflect = Reflect(model=OpenRouterModel())
out = reflect.run(goal="Define an LLM for a teen", answer="It's an AI thing.")
print(out.critique, out.revised, sep="\n---\n")

---

Examples (full case studies)

Each example is a complete, runnable agent. Default backend is OpenRouter. Pass --provider ollama to run locally instead.

export PYTHONPATH=src
export OPENROUTER_API_KEY="sk-or-..."

Example	What it shows	Run it
Research Assistant	Tool use, multi-step reasoning, report generation	python examples/research_assistant/run.py
Email Approval	requires_approval, checkpoint gates, side-effect policy	python examples/email_approval/run.py
File Patcher	Side-effect staging (propose allowed, apply blocked)	python examples/file_patcher/run.py
Ticket Triage	Multi-step classification, budget tracking across calls	python examples/ticket_triage/run.py

---

Architecture

┌─────────────────────────────────────────────────┐
│                   bear.run(task)                │
├─────────────────────────────────────────────────┤
│                                                 │
│   Task ──► System Prompt + User Message         │
│                     │                           │
│                     ▼                           │
│              ┌─────────────┐                    │
│              │  Run Loop   │◄── Budget check    │
│              └──────┬──────┘    each step       │
│                     │                           │
│          ┌──────────┴──────────┐                │
│          ▼                     ▼                │
│     Text response         Tool calls            │
│     ──► Report            ──► Policy check      │
│                               │                 │
│                    ┌──────────┼──────────┐      │
│                    ▼          ▼          ▼      │
│                 Allowed    Blocked    Approval  │
│                 ──► Execute ──► Skip  ──► Pause │
│                     │                    │      │
│                     ▼                    ▼      │
│                Tool result          Checkpoint  │
│                ──► Loop              (resume)   │
│                                                 │
├─────────────────────────────────────────────────┤
│  State (explicit, snapshotable, diffable)       │
│  Budget (steps, tokens, cost — all tracked)     │
│  Report (full receipt of everything)            │
└─────────────────────────────────────────────────┘

See docs/architecture.md for the full breakdown.

---

Philosophy

Most frameworks optimise for demos. BareBear optimises for understanding.

Read the manifesto.

---

Teaching with barebear

If you teach a course or club using barebear, we'd love to know. Two ways to tell us:

• Star the repo so other teachers can find it.
• Open an issue with the tag taught-with.

Lesson contributions are welcomed — see CONTRIBUTING.md.

---

Contributing

PRs, bug reports, lesson contributions, and honest feedback all welcome. See CONTRIBUTING.md.

---

Licence

MIT — Richey Malhotra, 2026.