01 · HARNESSES & WORKFLOWS TESTED PRINCIPLES
Design the environment
around the model.
The product is the loop: context, tools, handoffs, memory, checkpoints, and the interface that makes each step legible.HUMAN CONTROL · STEER · APPROVE · INSPECT · RECOVER
WHY THIS CHANGES THE PRODUCT
A polished chat surface cannot repair a weak loop.
Context should be selected for the next decision, not accumulated without limit. Tool interfaces are product interfaces: names, parameters, responses, and errors shape agent behavior. Handoffs must carry the request, state, evidence, restrictions, and unresolved decision, not merely a summary.
Memory, progress, and permissions are separate system concerns. If the interface collapses them into a confident answer, it hides the product's real state.
FIVE TOPICS
The loop is the product
What happens between intent and completion?
Context is a decision system
What must the agent know now?
Tools are contracts
Can the agent call and recover without guessing?
Handoffs preserve ownership
What survives when work moves?
Memory is not a transcript
What should persist, expire, or be rechecked?
PATTERNS TO USE
Observable agent loop
Every step has an owner, state, and stop condition.
RELEASE CHECKContext budget
Freshness, relevance, and disclosure limits stay visible.
RELEASE CHECKTool contract
Inputs, outputs, errors, authority, and effects are explicit.
RELEASE CHECKHandoff packet
Goal, owner, state, evidence, restrictions, and acceptance travel together.
RELEASE CHECKResume after interruption
Read current state before repeating an external effect.
RELEASE CHECKAsk for the missing decision
Research first; ask only where a person must choose.
RELEASE CHECKSource-of-truth refresh
Recheck action-driving facts at the system that owns them.
RELEASE CHECK“Do not design the agent's answer. Design the conditions under which it can keep working without losing truth, control, or the next decision.”Ayodhya's take
FAILURE MODES
| More context, less relevance | Stale or contradictory action |
|---|---|
| Tool name hides effect | Wrong action or unsafe approval |
| Summary-only handoff | Lost restrictions and evidence |
| Progress inferred from narration | Claims work that did not land |
| Retry without idempotency | Duplicate external effect |
| Memory treated as authority | Habit becomes unauthorized action |
PRIMARY SOURCES · REVIEWED 16 SEP 2026
AnthropicEffective context engineering for AI agents↗AnthropicWriting effective tools for agents↗OpenAIHarness engineering↗OpenAIMCP and Connectors↗OpenAIAgents SDK handoffs↗