Practical notes: When Should an Agent Loop Stop? DeepSeek Harness and Pi Give

The following notes reconstruct a practical path around “When Should an Agent Loop Stop? DeepSeek Harness and Pi Give Two Completely Different Answers”. Emphasis stays on contracts, checks, and drop-in code placeholders rather than motivational framing. When working through the Overview stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Opening: the question nobody asks until it breaks something

The Opening the question nobody stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

504 Gateway Timeout

Part 1 — What “stopping” actually is (it’s not one thing)

The Part 1 What stopping stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Model Request
    ↓
Assistant Response
    ↓
Tool Calls
    ↓
Tool Results
    ↓
Model reviews results
    ↓
Turn End
    ↓
Agent Idle
    ↓
Follow-up / Goal / Scheduler
    ↓
Next Turn

Part 2 — Why this matters now, not later

The Part 2 Why this stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve. The Part 2 Why this stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Part 3 — How it actually works: taking the machine apart

For the Part 3 How it stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Assistant:
"I've created PR #1234."

model stopped   ≠   turn stopped

3.2 Layer two: the tool batch finishes

For the 3 2 Layer two stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

# Illustrative example — conceptual pseudocode, not an official API
while has_more_tool_calls:
    response = call_model()
    tools = parse_tool_calls(response)
    execute_tools(tools)

Model
  ↓
create_pr()
  ↓
200 OK
  ↓
PR #1234

Goal
 └─ Goal Round
     └─ Turn
         ├─ Step
         │   ├─ Model Request
         │   └─ Tool Calls
         ├─ Step
         │   ├─ Model Request
         │   └─ Tool Calls
         └─ ...

3.3 So when does a turn actually end?

For the 3 3 So when stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary. For the 3 3 So when stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

3.4 The boolean trap: shouldContinue()

When working through the 3 4 The boolean stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Plugin A → false
Plugin B → true
Plugin C → false

shouldContinue(): boolean

3.5 The cost of this design

When working through the 3 5 The cost stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Prompt → Search → Answer

3.6 Pi’s answer: keep the loop small, split “continue” into two APIs

When working through the 3 6 Pi s stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours. When working through the 3 6 Pi s stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

// Illustrative example — conceptual pseudocode, not an official API
while (true) {
  while (hasMoreToolCalls || pendingMessages.length > 0) {
    callModel();
    executeTools();
  }

const followUps = await getFollowUpMessages();
  if (followUps.length > 0) {
    pendingMessages = followUps;
    continue;
  }
  break;
}

Inner Loop
    ↓
More tools? More steering?
    ↓
No
    ↓
Outer Loop
    ↓
Any follow-up?
    ↓
No
    ↓
Agent End

Steering  = you are changing my direction right now
Follow-up = you were about to let me stop, but I'm giving you one more thing

3.7 Laying DSH and Pi side by side

The 3 7 Laying DSH stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

DSH                              Pi
Goal                             Outer Loop
  ↓                                 │
Goal Round                    Follow-up Queue
  ↓                                 │
Turn                                ▼
  ↓                            Inner Loop
Step                                │
  ↓                        Steering + Tool Calls
Model ↔ Tools ↔ Inbox               │
  ↓                                 ▼
Turn-Stopping                   turn_end
  ↓                                 │
Turn End                            ▼
                                agent_end

3.8 Pi’s quieter but very useful addition: shouldStopAfterTurn

The 3 8 Pi s stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

shouldStopAfterTurn()   ≠   killEverything()

Agent
  ↓
fix bug
  ↓
run tests
  ↓
tests pass
  ↓
turn_end
  ↓
shouldStopAfterTurn = true
  ↓
agent_end

3.9 DSH’s more state-machine-like posture

The 3 9 DSH s stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve. The 3 9 DSH s stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

3.10 Why “why it stopped” matters more than it sounds

For the 3 10 Why why stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Agent stopped.

Agent stopped because:
- task completed?
- user cancelled?
- token limit hit?
- tool error?
- process crashed?
- blocked by policy?
- waiting on a human?

3.11 Pi deliberately doesn’t push all of this into core

For the 3 11 Pi deliberately stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

3.12 The price of that choice

For the 3 12 The price stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary. For the 3 12 The price stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

3.13 Tools that get a vote: concludesTurn

When working through the 3 13 Tools that stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

3.14 Even the AND/OR choice encodes a design philosophy

When working through the 3 14 Even the stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Tool A → concludesTurn = true
Tool B → false
Tool C → false

3.15 Concurrency adds yet another layer

When working through the 3 15 Concurrency adds stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours. When working through the 3 15 Concurrency adds stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

A = create PR
B = inspect dependency graph
C = run security scan

3.16 max-tokens: the textbook counterexample to "just retry"

The 3 16 max-tokens the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Model:
"I'm about to call deploy_production with
{ ... half-finished JSON ...

3.17 The read-only vs. side-effecting split

The 3 17 The read-only stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Part 4 — What most people get wrong

The Part 4 What most stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve. The Part 4 What most stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

DISCOVERED → MODIFIED → TESTING → PR_CREATED → CI_RUNNING →
WAITING_REVIEW → CHANGES_REQUESTED → MERGED → DEPLOYING → OBSERVING → VERIFIED

Part 5 — the judgment

For the Part 5 the judgment stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Part 6 — What this actually changes in practice

For the Part 6 What this stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Current side-effect state: UNKNOWN
Can it be resolved against external facts? YES
  → query the PR
  → query the branch
  → query the commit
  → query remote state

PR exists  → continue the existing workflow
PR absent  → retry creation

Goal:
  all v1 imports removed
  AND unit tests pass
  AND integration tests pass
  AND PR created

Conclusion: the sentence worth keeping

For the Conclusion the sentence worth stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary. For the Conclusion the sentence worth stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

If you’d like to show your appreciation, you can support me through:

When working through the If you d like stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Operational checklist

For the Operational checklist stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state.

Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node.

Pin dependency versions and record the image digest that ran the demo. Reproducibility beats tribal knowledge.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for a95f0d36f035: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.

For the hardening note 0 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Hardening detail 0/816: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 1 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Hardening detail 1/816: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 2 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Hardening detail 2/816: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 3 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 3/816: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.