Work with agents

Agents are a first-class interface to Sentient-X. The project connection installs a shared skill and MCP server; repository agents can also inspect code, run tests, and implement bounded changes. The best results come from specifying the outcome, boundaries, and proof.

The four-part brief

Part	Tell the agent	Example
Outcome	What should become true	“Make this episode discoverable for training.”
Scope	The project, task, member, or failure in bounds	“Dataset `piper-cups`, task revision X.”
Boundaries	What requires approval or must not change	“No raw deletion; no physical run; preserve wire types.”
Evidence	What proves completion	“Receipt, lineage query, focused test, then relevant gate.”

Ask for action, not theater

“Inspect, implement, test, and hand back evidence” is stronger than “write a plan.” Ask for a plan only when a real choice would change scope, safety, or architecture.

A reliable agent loop

Orient

Have the agent read canonical guidance and inspect current state. For platform work, use MCP reads; for repository work, trace the real production entry point and tests.

Bound

Name spend, hardware, held-out evaluation, raw-data, compatibility, and deployment constraints before mutation.

Act

Let the agent complete routine, reversible work inside the boundary. Avoid prescribing files before it has found the owning abstraction.

Verify

Require typed receipts, exact resource identities, focused checks, and the broadest relevant gate. Let refusals stay refusals.

Handoff

Ask for the outcome first, then changed files or resources, proof, remaining risk, and the next safe action.

Prompts that scale with experience

Teach me this subsystem by tracing one real request from entry point to effect and test.
Name the canonical types and what can fail. Do not edit anything. Cite every file.

Make <outcome> true in <scope>. Read the applicable guidance first. Preserve <invariants>.
Find the smallest vertical change, implement it, run focused checks, then the broadest
relevant gate. Do not touch unrelated user changes. Hand back evidence and remaining risk.

Using the configured Sentient-X project, inspect <resource>. If the next action is reversible
and within <budget>, execute it and watch the durable operation to a terminal state.
Ask before physical motion, promotion, cancellation, or wider spend. Cite every receipt.

Keep these decisions human

widening task or project intent;
spending outside the stated budget;
admitting hardware or starting physical motion;
promoting to production or rolling back a healthy deployment;
weakening safety, provenance, held-out evaluation, or raw-data guarantees.

Agents can assemble the evidence and recommend an action. The accountable engineer owns these decisions.

Agents share the public contract

The hosted Improve operator and a locally connected coding agent call the same project-scoped pillar doors. Missing capability belongs at the owning API/MCP service, not in a private agent implementation.