Agent Capabilities
Capability mapping prevents over-promising and guides tool design and evaluation.
Capabilities are the job description skill section — list what the hire can actually do.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Capability Tier Ladder
Scroll inside the frame to explore · use + / − to zoom up to 200%
Start at L1 and only climb when evals prove reliability at each level.
Key Takeaways
- 1.Capabilities = concrete skills not vague AI magic.
- 2.L3 autonomous planning with recovery.
- 3.L4 multi-agent long-horizon tasks.
Real Example
Scenario
L2 data analyst agent: NL→SQL, matplotlib chart, Slack post. It does not auto-deploy models or open PRs — that is L4 multi-agent scope.
What you would do
Capability map: L2 = multi-tool loop (SQL → chart → Slack). It lacks L3 re-planning when SQL returns zero rows and lacks L4 code-review agents. Evaluate with 10 analyst questions and task-completion rate, not SWE-bench.
Practice Task
Rate your capstone idea L1–L4 using the tier ladder in the diagram. List the one tier you must nail before climbing to the next.
Code Walkthrough
Highlighted lines show where Agent Capabilities happens in the code.
1CAPABILITY_TIERS = {2 "L1": ["calculator", "weather"],3 "L2": ["run_sql", "create_chart", "slack_post"],4 "L3": ["planner", "run_sql", "create_chart", "replan_on_empty"],5 "L4": ["supervisor", "analyst_agent", "reviewer_agent"],6}7
8def tools_for_tier(tier: str) -> list[str]: # define a reusable function9 return CAPABILITY_TIERS[tier] # return the result10
11analyst_tools = tools_for_tier("L2") # key line for Agent Capabilities12print("L2 analyst gets:", analyst_tools) # not L4 deploy/PR toolsCheat Sheet
Quick recap
quick ref- •Map capabilities before building
- •Match evals to each tier
- •L1 = one shot tool
- •L3 = plan and recover
- •L4 = multi-agent horizon
Common Mistakes
- ✕Skipping evaluation for Agent Capabilities before production
- ✕No logging or tracing around agent capabilities steps
- ✕Ignoring cost and latency implications