Launch — March 2026 · v0.1.1
BareBear

Every step leaves a print.

A minimal, policy-first Python agent framework.

Richey Malhotra · MIT Licensed

The Problem

Most agents fail the same way in production

👻
Hidden state no one can inspect
Tool calls that fire without gates
♾️
Unbounded loops with no kill switch
No answer to "why did it do that?"

"Hidden memory is technical debt with compound interest."

Raison d'être

Make agent behaviour governable

01
Explicit state
02
Policy before action
03
Bounded autonomy
04
Run receipts

Strong bones. Nothing extra.

The Bones

7 primitives.
Nothing extra.

BearThe agent runtime
TaskA unit of work
StateExplicit, diffable, snapshotable
ToolGated callable with risk metadata
PolicyPre-execution constraint set
CheckpointHuman approval gate
ReportFull run receipt
Policy Engine

Policy checks before execution

Because you can't un-send an email.

Tool call requested
?
Blocked? Side effects? Approval?
✓ Allow
Execute
✗ Block
Skip
⏸ Pause
Checkpoint

"Tools are permissions, not features."

Bounded Autonomy

An agent without limits is a while-True loop with a credit card

Steps 7 / 10
🔤
Tokens 8,200 / 50,000
💲
Cost $0.04 / $0.10
🛑
send_email → approval required
Run Receipts

Every run is a receipt

BareBear
Run receipt · policy-first execution log
Summary
task_ida1b2c3d4
statuscompleted
steps5
tokens8,200
cost$0.0340
duration1.34s
Assumptions
Tier assumed from email domain
Uncertainties
Could not verify account status
Every step leaves a print.

"I don't know" is more valuable than a confident hallucination.

Landscape

Different tools,
different jobs

No orchestration theatre here.

Category
Optimised for
Orchestration-first
e.g. LangChain
Breadth & composability
Vendor-native stacks
e.g. OpenAI Agents
Fast onboarding, ecosystem lock-in
Autonomous systems
e.g. OpenClaw
High autonomy, broad tasking
BareBear
Explicit, policy-first, auditable
Is It For You?

Choose BareBear if…

"Power should be earned, not assumed."

Now Live
BareBear

Built for agents you can explain after they run.

MIT licensed · Python 3.9+ · 83 tests
github.com/richey-malhotra/barebear

What would you need before trusting an agent in production?

Richey Malhotra · richey-malhotra