Provider path
OpenCode Go supplies the fixed DeepSeek V4 Flash model. The teaching contract stays separate so model behaviour can be tested without giving the model authority.
EXPERIMENTAL WORKSHOP · 01Build the teacher
Build the teacher
behind the lesson.
A real, inspectable agent that teaches agent engineering. Ask it a question, watch it retrieve a bounded source through MCP, and then trace which layer made the answer possible.
THE STACK
One lesson, four distinct responsibilities.
Application loop
The SDK owns the model call, typed tool set, step budget, and the response boundary the browser can inspect.
Capability boundary
The curriculum server exposes read-only tools. The agent receives only the allowlisted capabilities it can use in this lab.
Durable shape
The separate Eve workspace makes instructions, tools, connections, skills, and evals first-class files instead of hidden glue.
OPEN THE AGENT
The filesystem is part of the lesson.
Run the browser adapter below, then inspect the durable Eve version in agent/academy-teacher.
agent/instructions.mdThe always-on teaching contractagent/agent.tsEve runtime and model configurationagent/connections/academy-curriculum.tsMCP connection boundaryagent/tools/inspect-teaching-move.tsTyped local Eve toolevals/teaching-contract.eval.tsRegression eval casesRUN THE LOOP
PUBLIC LEARNING LAB · NO SERVER MEMORYAsk the teaching agent
This public learning lab sends your question to DeepSeek V4 Flash through OpenCode Go and AI SDK. Academy gives the loop one read-only curriculum-tool execution through MCP and does not save the question or reply on its server. The provider may process the content under its own retention policy. Do not enter personal, confidential, customer, credential, or production data.
BREAK THE CONSENSUS
DETERMINISTIC · ZERO MODEL CALLSEvidence-Invariant Mutation Harness
This test deliberately feeds an agent team cases where confident agreement is wrong. Fixed assertions—not another model—decide whether the evaluator catches the failure.
Remove one decisive fact. The safe answer must become uncertain.
Change current approval to draft, stale, or revoked. The action must follow authority.
Give one role the only sourced objection. Nine votes cannot erase it.
Reorder irrelevant details. The evidence-backed decision must stay stable.
12 frozen cases4 calibrated controls · 4 unanimous-wrong traps · 4 transfer cases
Remove the MCP tool and predict what changes.
Then add a new read-only tool, write the regression case first, and compare the trace. The point is not to make the agent more autonomous; it is to make every new capability visible, bounded, and testable.
Continue with the MCP chapter →