Building agents with code.
agent-lab scenarios
13 folders · 5 files

README.md

Scenarios #

SDK-agnostic “what to accomplish” docs. Each scenario folder has a README.md (Goal / Success / Assert / Implementations) and optional <framework>/index.ts demos.

Status legend #

Badge Meaning
verified Demo path + tripwire PASS (AGENT_LAB_RESULT ok:true) — smoke-backed
linked — pending keys+PASS Demo path present; not done until keys + PASS (SKIP ≠ PASS)
gap / stub Nothing to link, or note-only

Source of truth: GAPS.md. linked ≠ verified — a linked OA / CAS / Managed cell is pending, not complete.

Unit 1 teaching spine (CTA) #

Start here → finish at Support. Teach the smallest useful TypeScript path to a dependable support agent (Letta verified on-ramp first). Employer Managed platforms are secondary explorations, not the lead.

Done definition (shared): Support green on Letta (support/letta tripwire PASS) is Unit 1 done. Citation is optional after Support; computer-use is advanced. Linked OA / CAS / Managed cells are compare after Unit 1 (linked ≠ verified) — not alternate finishes. Soft-pedal Managed.

  1. Phase-1 on-ramp — Letta verified blocks: hello → tool-use → streaming → memory → approvals (TS demos; tripwire PASS).
  2. Reply Obligation: preserve the originating thread/route and delivery receipt across a context wipe; this is the first graduation pattern.
  3. Support ← CTA end / done: retrieve from a stub KB, then explicitly resolve with grounding or escalate when the KB cannot answer.
  4. Citation grounding (optional): cite-or-refuse from stub sources — stricter than Support; optional after Support. See citation-grounding/.

Path: UNIT-1.md. Facilitators: UNIT-1-WORKSHOP.md. Playlist: tripwire --playlist unit-1-letta. OA / CAS / Managed paths for the same scenarios stay linked — pending keys+PASS — compare after Unit 1, not alternate finishes. Coverage: GAPS.md. Computer-use remains advanced after citation, not a Unit-1 graduation step.

Scenario One-liner
hello One prompt → one short reply
tool-use Call a trivial add tool for 41 + 1
streaming Observe incremental turn events
memory Retain state across reconnect / second turn
approvals Gate a sensitive action on human/host approval
reply-obligation Graduation — deliver back to the originating thread after a context wipe
support Graduation — KB-grounded resolve or escalate
citation-grounding Graduation (optional) — cite-or-refuse from stub sources
handoffs Router delegates to a specialist
mcp MCP tool server — CAS in-process echo linked (other stacks gap)
sandbox-files Sandbox file read/write — CAS local workspace Write+Read linked (other stacks gap)
skills Skill packages — CAS filesystem lab-marker linked (other stacks gap)
computer-use Advanced — GUI/browser/Computer adapter + observable artifact

Employer Managed / platform adapters (Bedrock, hosted enterprise runtimes, etc.) stay secondary — same evidence bar (verified only after PASS); do not invent client APIs.

Coverage: GAPS.md. SDK catalog: ../sdks/.