Skip to main content

5th Grade Summary

Not every AI task should become an agentic workflow.

Normal chat should stay fast and use the right route and data.

KamAgentic should handle bounded internal work like preparing labels, promoting fixtures, and building release packets.

The split keeps user answers fast and operations safer.

Agentic systems are powerful when the task is long-running, stateful, and approval-based.

They are risky when used as a replacement for normal product routing.

Kam's architecture needs both modes. Chat should answer through route contracts and hot reads. KamAgentic should run bounded internal workflows where the system can pause, resume, gather evidence, ask for approval, and create durable artifacts.

The split

Chat path vs agentic path

Question
User asks why a line moved
Normal chat
Route to market-shape answer
KamAgentic
Not needed unless follow-up creates work
Question
User asks which games counted
Normal chat
Load denominator hot read
KamAgentic
Not needed
Question
Failed trace needs labeling
Normal chat
Not a user chat task
KamAgentic
Draft label review packet
Question
Approved label needs fixture
Normal chat
Not a user chat task
KamAgentic
Create fixture candidate
Question
Release needs evidence packet
Normal chat
Not a user chat task
KamAgentic
Build release packet
Question
Long-running data repair
Normal chat
Backend or infra workflow
KamAgentic
Only if exposed as bounded work item
Question
Human approval needed
Normal chat
Show answer caveat
KamAgentic
Pause and wait for KamOps approval

Takeaway: Agentic work belongs where state and approval matter. Chat belongs where route contracts and hot reads can answer directly.

What KamAgentic should run

Good KamAgentic workloads include:

  • agentic.trace_label_prep.v1
  • agentic.fixture_promotion.v1
  • agentic.release_packet.v1
  • agentic.review_packet_builder.v1
  • agentic.scorecard_summary.v1
  • agentic.drift_investigation_prep.v1

These workloads are bounded. They have inputs, outputs, states, allowed tools, artifacts, and approval points.

Workflow

Agentic trace-label prep

Kam should help the user move from a question to evidence, caveat, decision, result, and review.

  1. 1

    Read failed trace

  2. 2

    Draft expected intent

  3. 3

    Draft entities and route

  4. 4

    Run deterministic graders

  5. 5

    Prepare review packet

  6. 6

    Pause for human approval

  7. 7

    Store approved label

  8. 8

    Create fixture candidate

The agent prepares evidence. The human approves truth.

Durable state matters

An agentic workflow should survive retries and interruptions.

That means each run needs:

  • run ID
  • workload ID
  • input artifact URIs
  • current state
  • step history
  • generated artifacts
  • approval requirements
  • idempotency keys
  • final status

Trust receipt

What Kam should prove before confidence

A useful answer should leave a small receipt: route, scope, freshness, evidence, missing data, and confidence state.

Route

agentic.fixture_promotion.v1

Scope

An internal workflow that turns an approved label and failed trace into a fixture candidate.

Freshness

The run is waiting for human review before the fixture can become release-gated.

Evidence loaded

  • approved label loaded
  • failed trace loaded
  • fixture candidate generated with grader set

Missing or caveated

  • human approve/edit/reject decision
  • release gate attachment
Status: waiting_for_review

Why not Temporal now

Temporal is a strong durable execution platform.

Kam may want it later if workflow scale, retries, visibility, or long-running orchestration become too complex for the current adapter.

But adopting Temporal too early would add operational weight before the workload boundaries are fully settled. For now, a custom runAgenticWorkload(...) path with DynamoDB/S3 artifacts can be enough if it stays disciplined.

The standard should be:

  • bounded workload IDs
  • explicit states
  • compact indexes
  • durable artifacts
  • human approval gates
  • idempotent writes
  • clear failure receipts

Agentic adoption rules

Bounded

Every workflow needs a declared input, output, state machine, and allowed tool set.

Auditable

Every run should create receipts, artifacts, and reviewable decisions.

Approval-aware

Human approval should be a first-class state, not an afterthought.

Takeaway: The right agentic framework is smaller than general autonomy and stronger than a script.

The lesson

The better Kam framework does not make everything agentic.

It makes the right things agentic.

User-facing chat should stay fast, grounded, and route-contract-first. Internal AI operations should use KamAgentic when the work is multi-step, stateful, evidence-heavy, and approval-based.

The next action is to make agentic.trace_label_prep.v1 the first polished workload and use it to feed KamOps label review.

Read next

Related field notes

View all posts
kam-frameworkai-operating-system

Kam AI Is Becoming an Operating System

Why Kam is moving from AI chat into a production loop of traces, labels, graders, fixtures, release gates, and agentic work.

9 min read

kamopshuman-labeling

Human Labeling Is the Center of KamOps

How KamOps turns failed traces into approved labels, fixtures, and review packets without letting automation become the source of truth.

8 min read

kam-frameworklessons-learned

Lessons From Building the Better Kam Framework

What Kam learned moving toward a framework of KamSRE, KamOps, KamEvals, KamAgentic, labels, fixtures, and workload scorecards.

10 min read