Plan→ Act→ Deliver↻
Ask aibuddy.Get it done.
Describe the outcome. aibuddy plans the steps, uses tools, checks the result, and hands back something you can use.
From request to deliverable
A real run
Find the salary range for production and technical leads at medical-device MNCs in Shanghai and Suzhou.
Understands the outcome›
Breaks the work into 5 steps›
Searches Shanghai salary sources›
Searches Suzhou salary sources›
Cleans data and computes ranges›
Checks against saved baselines›
Writes the final report›
Checks the answer before delivery›
Report ready — 7+ sources checked, two cities compared, and assumptions called out.
salary-distribution.md
9 sections · 7+ sources · 13.5 KB
Tools for real work
Planning · search · sandbox …
Planning
Turns a vague request into ordered, checkable steps
Ask User
Asks when a decision needs you instead of guessing
Web Search
Finds fresh public sources and keeps citations visible
Shell
Runs scripts in isolation and inspects real output
Memory
Carries useful facts and preferences into future work
Subagent
Splits independent work across parallel agents
Also: Decision gates · Knowledge · Skills · Agent teams · Kanban · Wait & resume · MCP
Parallel when it helps
Task-dependency DAG
Parallel · no deps
T1Search Shanghai
T2Search Suzhou
T3 needs T1 · T2
T3Verify sources
T4 needs T3
T4Synthesize findings
T5 needs T4
T5Write the report
Team · 5 agents
Researcher1
Research · SH
Researcher2
Research · SZ
FactChecker
Verify
Synthesizer
Synthesize
Writer
Write
→ merged into 1 report
Measured, not guessed
SWE-bench · GAIA · τ-bench …
0
Agentic score
$0.00
Cost / run
0×
Compaction
SWE-bench Verified0+9
GAIA 0+6
τ-bench retail0+5
WebArena 0+7
OSWorld 0+4
Start with the outcome
Tell aibuddy what finished looks like. It works backward from there.
A usable report from a question
Give it a topic. It searches, cross-checks sources, and shapes the findings into a file you can read, share, or keep editing.
examplecompetitor research → report.md
Clean data and a script
Hand it raw tables, logs, or a code goal. It writes scripts, runs them, checks the output, and fixes what fails.
examplemessy CSV → cleaned.csv + script
A finished browser workflow
When work lives behind a login, it can use your browser with permission, take visible actions, and leave you with the export or update.
exampleadmin console → batch export
A long task moved to done
Multi-step work can keep moving without being watched the whole time, with progress you can inspect, pause, and resume.
example20-step workflow → completion log
Capabilities that finish the work
From prompt to verified delivery
Most AI tools stop at generating an answer. aibuddy is built for the work after that: planning, acting in real environments, coordinating specialists, preserving state, and checking whether the outcome is actually complete.
Field notes from building aibuddy
What we are learning about agents, context, evaluation, and AI-native work.