Building agentic workflows on Claude Fable 5.

Most teams who tried agentic AI workflows in 2024 and 2025 gave up because the model could not reliably hold state across more than three or four steps. Fable 5 changes that boundary. Carrying twelve to twenty sequential decisions through a coherent output is now realistic. But the model is one layer. The architecture around it is where production reliability comes from.

The five layers of a production agentic workflow

Every agentic loop IG ships in production has the same five layers. Skipping any of them causes specific failure modes.

Layer 1: State management

Long-horizon workflows need explicit state. The 1M context window of Fable 5 makes it tempting to throw everything into the prompt and let the model figure it out. This works for prototypes. It breaks in production when conversations get long enough that important state gets buried.

The pattern that works: maintain a structured state object outside the model that gets serialized into the prompt at each step. Pass the state through, not just the conversation history. The state object tracks what has been decided, what is pending, what the current goal is, and which tools have been invoked.

Layer 2: Tool integration

Fable 5 is strongest when it has access to real tools. The model knows how to use a calendar API, a database, a search engine, or a file system far better than any prior public model. The architectural decision is which tools to expose.

The rule of thumb: expose the smallest set of tools that lets the workflow complete. Each additional tool increases the chance of off-path behavior. Most production agentic workflows need 4-7 tools, not 30.

Layer 3: Fallback handling

Fable 5 routes roughly 5% of sensitive-domain queries to Opus 4.8 automatically. For most consumer-facing workflows this is invisible. For agentic workflows that touch regulated content or sensitive analysis, the fallback can mid-task switch the model. Your architecture needs to handle this gracefully.

The pattern: design the prompt and state schema so they work identically on Fable 5 and Opus 4.8. Test the workflow against both models during development. Log which model handled each step so you can audit behavior after the fact.

Layer 4: Human-in-the-loop checkpoints

Production agentic workflows have human checkpoints. Even with Fable 5’s improved reliability, irreversible actions (sending external communications, executing transactions, modifying production data) should pause for human approval. The IG Next Best Action product is built on this principle: the system surfaces a recommendation, the human approves or modifies, the system executes.

The architectural pattern: insert checkpoint nodes at every step that has external consequences. The checkpoint can be a Slack message, a UI confirmation, or a queued review depending on the latency budget. Build the checkpoint into the workflow graph, not as an afterthought.

Layer 5: Observability

You cannot debug an agentic workflow you cannot see. Every step of every loop should emit a trace: which prompt was sent, which model responded, what tools were called, how long each step took, what the state looked like before and after. The observability layer is non-negotiable for production agentic systems.

The pattern IG deploys: structured logging at every step, sampled to 100% during the first month in production, then sampled to 10-20% once the workflow is stable. Pair the logs with a small dashboard that surfaces failure modes, latency outliers, and unexpected tool-use patterns.

The reference architecture

Most production agentic workflows IG ships fit a reference architecture with these components:

1. Trigger. Webhook, schedule, or user action that starts the loop.
2. State initializer. Creates the workflow state object and loads any necessary context (customer data, historical state, business rules).
3. Step runner. Calls Fable 5 with the prompt, state, and tool definitions. Handles the response. Updates the state. Determines the next step.
4. Tool executor. Mediates between the model and the external systems. Implements rate limiting, authentication, and error handling.
5. Checkpoint manager. Pauses the workflow at human-in-the-loop nodes and resumes once approval lands.
6. Observability layer. Logs every step, every tool call, every state transition.
7. Outcome handler. Writes the workflow result to its destination and emits success or failure signals.

The patterns that go wrong

Three failure modes show up often enough to flag.

The "everything in the prompt" anti-pattern. Loading all historical context into every prompt seems efficient until the conversation gets long enough that important early state gets diluted. State management via a structured object is the fix.

The "no human in the loop" anti-pattern. Teams confident in the model’s capability remove human checkpoints. Six weeks later, a single off-path output causes a real-world consequence and the workflow gets shut down. Checkpoints at irreversible action boundaries are the right default.

The "no observability" anti-pattern. When the workflow fails, the team cannot reconstruct what happened. Cannot fix the bug. Cannot improve the prompt. Cannot calculate the per-call cost. Observability built in from day one pays back every single week.

Cost engineering inside agentic workflows

Token costs add up fast in agentic workflows because every step carries forward state. A 20-step agentic loop with a state object that grows over time can easily hit 50,000-200,000 tokens per execution. At Fable 5 pricing ($10 input / $50 output per million tokens), that is meaningful spend per workflow run.

Three cost engineering moves that compound:

State pruning. Each step prunes the state object to what the next step actually needs. Aggressive pruning typically reduces total token spend by 30-50% with no loss in workflow quality.

Tier-aware routing inside the workflow. Not every step in an agentic workflow needs Fable 5. Steps that are simple transformations, classifications, or extractions can route to Haiku 4.5 and feed their output back into the Fable 5-led main loop. This pattern reduces workflow cost by 40-70% in most deployments.

Batch processing where possible. Anthropic offers 50% off batch pricing on Fable 5 ($5 / $25 instead of $10 / $50). For non-real-time agentic workflows, queueing and batching is a clean cost win.