## Session Log

### Prepared flow

_Fill this in before review day._

**Test Environments:** If evaluating your agent requires setting up an environment
(such as e.g. setting up a git repo to operate in, for GitBot from the [lecture](../../lectures/03-agents.md)),
you need to set up 3-5 such environments that can be used during the review session.
Some agents (e.g. a research assistant) might not require a test environment;
then you can skip this step.

**Demo Scenario:** Pick one environment from the set above and
describe a task you want the agent to accomplish in this scenario.

**Agent trace:** Paste the trace of agent behavior on your demo scenario;
the trace should include user and assistant messages,
as well as the tool calls that the agent makes and their outputs
(if a tool has a textual output).

**Outcome:** _What was the final result? (screenshot or describe)_

**Guardrails:** Describe which guardrails your agent implements
(what are the actions that are disallowed or require user approval)
and illustrate those guardrails on your demo scenario.


---

### Live session

- Demo your agent using the demo scenario from above.
- Tell the reviewers what other test environments are available (if applicable)
- Allow each reviewer to interact with your agent in a test environment of their choosing.
  You can drive your demoing laptop as the project author,
  but take the prompts from them.

The project author should fill in the sections below in the same format as
above, during the session.

#### Reviewer 1

**Reviewer name:**

- What was the task?
- What was the end result (what did the agent accomplish)?
- Which tools did the agent call?
- What happened when you asked for an action that is not automatically allowed by the guardrails?

---

#### Reviewer 2

_Copy the structure above for additional reviewers._

---

### Full group

After all reviewers have interacted with the agent, discuss and answer the following:

**Guardrail comparison:** _Across the test environments and reviewer prompts, how did the guardrails behave? Were there cases where the same guardrail blocked an action in one environment but allowed it in another? Were there actions that surprised reviewers (either blocked unexpectedly, or allowed when they expected a block)?_