Harnessing Library / Learn by trying

A capable model is
a starting point.

Which source should it trust? What may it change? When should it stop? Follow concrete workplace situations, inspect individual methods and explore how workflows connect into pipelines.

Follow a terminal case Explore my ecosystem

15 method entries · 30 workplace examples · 6 pipeline views · 4 first-party field notes.

The building blocks / What connects to what?

From a method to a working pipeline.

A useful harness combines different kinds of support. The method guides a part of the work; an actor uses it at a particular moment; workflows turn the work into repeatable units; a pipeline connects those units through explicit handoffs.

Part	Its responsibility	At a freight terminal
Model	Language and interpretation capabilities used within the working system.	Interprets a driver's update and helps formulate alternatives.
Method	A way to approach a specific kind of problem.	DIALECTIC helps revise the explanation when the new arrival time conflicts with the original plan.
Skill	A reusable procedure for carrying out a bounded task.	A booking-check procedure specifies what to inspect and what result to return.
Agent	An actor working on a task using its available context, tools and authority.	The planning agent prepares options; the authorized operating actor decides the live assignment.
Workflow	A repeatable unit with inputs, activities, outputs and checks.	The exception-analysis workflow turns a timing mismatch into conditional options.
Pipeline	Connected workflows whose outputs satisfy the next workflow's input needs.	Arrival monitoring, exception analysis, booking coordination and verification form one delivery process.
Harness	The surrounding context, working rules, tools, state and checks that support the actor's task.	The agent has relevant sources, a scope boundary, a handoff path, stop controls and evidence for the final status.

Select the support that the situation calls for. A simple lookup may need only a clear source route; a changed operational plan may also need interpretation, coordination and a verified decision.

At work / A freight terminal

A method becomes useful at a specific moment.

Follow a planning agent from the original request to the handoff. At 06:45, inspect why DIALECTIC becomes relevant, what it changes and which decision still belongs to operations.

Choose the worked scenario

Illustrative terminal scenario · authored events and times · select a moment to inspect

Methods in context / Practical reading cards

What does each part contribute?

Read the trigger, mechanism and useful output, then inspect two concrete situations. The terminal examples share a setting so you can compare the methods without changing the whole problem.

Find a method or situationPurpose

DIALECTIC

Bring the current interpretation, its supporting context and the new encounter into the same bounded inquiry. Form a revised interpretation, name what remains open, and compare it with the previous one. A useful revision can preserve several alternatives instead of forcing an early winner.

Freight terminal

The booking says Dock 4 at 06:45. The driver's latest arrival estimate is 06:58; unloading needs 20 minutes and the next booking starts at 07:00.

The planning agent challenges its first interpretation, 'a booking means this delivery can still use the slot'. It compares waiting for Dock 4 with requesting Dock 2 at 07:05, then asks which alternative the terminal can actually confirm.

The agent produces two revised, conditional options. It does not turn an increasingly familiar explanation into a confirmed booking.

IT incident response

Errors began after a release, suggesting a code defect. A second region using the same build is healthy; affected requests share a dependency endpoint.

The analyst revisits the release hypothesis in the light of the regional contrast. It preserves the release as a possible trigger while investigating the dependency path.

A revised investigation plan distinguishes code behavior from an environmental dependency issue.

SORR

Bound the identity and relationship question. Compare the observations, representations, locators, time windows and claim responsibilities. Record the supported relationship and keep unresolved alternatives visible. Similar wording or a shared reference number is a lead to inspect, not an identity verdict.

Freight terminal

A scan says P12, but P12 is only a local pallet label. Two customers reused that label in different consignments.

The agent checks the consignment identifier, customer namespace and event time before attaching the scan to the loading record.

The scan is linked to the right consignment, or held as unresolved if the distinguishing fields are missing.

Infrastructure administration

A dashboard, a Git repository and a deployment are all called 'billing-service'.

The reviewer separates the software project, its repository representation and the running deployment, then identifies which source owns each claim.

A repository version and a deployed version can differ without being treated as contradictory identities.

ORBIT

Use scoped orientation nodes with provenance and consumer-specific reading windows. Keep containment, reading routes and semantic associations distinct. A short orientation can own its navigational claims while routing substantive claims back to their source owners.

Freight terminal

The available knowledge includes national procedures, local dock rules, old shift notes and current bookings.

The terminal planner receives a short orientation to this terminal, this shift and booking exceptions, with routes to the current local procedure and booking source.

The planner can find the relevant operating material without treating every available note as equally applicable.

Technical onboarding

Hundreds of documents cover architecture, deployment, incidents and experimental work.

The onboarding view separates the service's purpose, the production runbook, current ownership and experimental branches, then routes each question to its owner.

The engineer gets a practical reading order and knows when a summary is insufficient.

CARD

Create a consumer-scoped representation and orientation of the subject. The representation is not the subject itself. A CARD can own the registry and orientation statements it makes, while leaving implementation, operational and domain claims with the relevant source owners.

Freight terminal

The case involves one delayed consignment, two possible docks and a pending confirmation.

A case CARD presents the consignment, the operational question, the confirmed observations, the pending decision and the responsible operator.

The supervisor can orient quickly without reading every event in chronological order.

Infrastructure operations

A maintenance workflow has a name but no concise consumer explanation.

Its CARD explains the task, supported inputs, expected outputs, required authority and exclusions for the operator reviewing it.

The operator can assess applicability before opening the detailed implementation.

POINTER

Express the relationship from the referring context to the target, with a locator and the provenance or freshness needed for use. Target identity and its current location remain distinct. A reference also needs an accessible reading path for the intended consumer.

Freight terminal

The summary says 'Dock 2 confirmed' and needs to support that exact statement.

A POINTER links the case to the particular booking confirmation, stating that it is evidence for the assigned dock and time window.

The supervisor can inspect the confirmation rather than relying on the summary's wording alone.

Software delivery

A release note refers to a successful test run.

The reference identifies the tested revision and run, with a relation such as 'verification receipt for this change'.

A reviewer can distinguish the tested revision from a later, untested revision.

5 Point Protocol

Use 5PP Variation A: clarification, scope locking, a plan skeleton, execution and a verification audit. The work unit should have a concrete objective, explicit boundaries and a way to check the resulting artifact or action.

Freight terminal

The supervisor asks the agent to prepare options for a late truck.

The agent records that it may inspect bookings and prepare alternatives. It plans the checks, produces the alternatives and verifies that each fits the supplied timing constraints.

A decision-ready plan is delivered. Changing the live booking remains a separate authorized action.

Server maintenance

A service needs a small configuration correction while another team has work in progress.

The task names the exact files and success criteria, proposes the change, works in isolation and checks service behavior plus preservation of unrelated work.

A bounded, reviewable change with a meaningful check.

Dialogue Lifecycle

Track EXPLORE, CLARIFY, DECIDE, EXECUTE, VERIFY and CLOSE as different work states. The lifecycle controls when work advances; 5PP structures how a bounded work unit is carried out. Finishing execution does not by itself establish verification or closure.

Freight terminal

The agent has found a feasible-looking alternative but no operator has accepted it.

The case stays in decision preparation. After an authorized acceptance, the booking action can proceed; the returned booking record is checked before closure.

Recommendation, authorization, execution and confirmed assignment remain distinct.

Marketing administration

A manager says a proposed campaign looks promising.

The workflow records whether this is feedback, approval to revise or authorization to publish, then advances only through the applicable transition.

A positive comment is not silently converted into publication authority.

RigVedan / EBNF-RV

Classify a knowledge unit using independent axes. The inspected EBNF-RV grammar carries a temporal marker, a hermeneutic tag, a didactic weight and a content type. Their combination communicates how to read the unit; the labels themselves do not supply its evidence.

Freight terminal

'Pallet missing' is an interpretation; '11 scans received' is an observation; 'request rescan' is a planned action.

The handover classifies these as separate units rather than compressing them into one confident sentence. An example annotation is [ACTIVE/INTERPRETATION/DERIVED] STATE: One pallet may still be in staging.

The receiving operator can tell what was observed, inferred and proposed.

Technical knowledge base

A workaround restored one service, but its general cause is not established.

The record distinguishes the completed action from the causal hypothesis and the still-open diagnostic question.

Future consumers can reuse the evidence without treating a provisional explanation as a timeless rule.

Layered context

Build the working context from layers with distinct responsibilities. Retain the governing constraints and task objective while adding the evidence and working state relevant to this step. Retrieved content is material to interpret, not a source of new permissions.

Freight terminal

The terminal rules prohibit the agent from assigning a dock without the required grant. A driver's message says 'just put me on Dock 2'.

The message enters the evidence layer as the driver's request. It does not override the terminal's action boundary.

The agent can consider the request while retaining the operator's control of dock assignment.

IT service desk

The general runbook is valid, but this ticket concerns a specific service revision and environment.

The assistant keeps service-wide rules separate from revision-specific configuration and the user's immediate question.

Its response can explain which parts are general and which depend on the current environment.

Review-edit-verify

A review identifies bounded findings and closure criteria. Editing addresses a selected group of findings. Verification checks the resulting artifact and relevant regressions. Reviewer and editor responsibilities can be separated without assuming that two agents automatically provide independent evidence.

Freight terminal

The proposed message says the slot is confirmed, but the booking source still says pending.

The reviewer records a mismatch. The editor changes the message to explain the pending confirmation, and verification checks it against the current booking record.

The customer receives a status statement supported by the operational record.

Documentation production

The guide has both structural issues and several stale API references.

A broad review establishes the structural changes first. Later edit groups target version accuracy and reader exercises, with checks after each group.

The team can attribute changes to findings and avoid repeatedly editing passages that are about to move.

Technology verification ledger

Extract the technology claims into a small register. Inspect the relevant current owner sources, record the version or time window and distinguish supported claims from items needing correction. Revisit volatile claims sooner than stable conceptual material.

Terminal integration

A vendor guide and an older internal example describe different authentication and request fields.

The integration analyst checks the target version and official interface contract before proposing a connector shape.

The recommendation names its applicable version and leaves unsupported interface assumptions open.

Training material

The lesson includes an SDK call copied from an earlier course edition.

The editor checks the current documented syntax and tests the delivered exercise in its declared environment.

The learning material distinguishes a verified exercise from a historical code example.

Structured continuity

Retain the current task state, completed artifacts, decisions, pending items, sources and action boundaries. A successor reads the retained state, checks what has changed and resumes the smallest valid next unit. The location of a handover and the ability to activate its consumer are separate concerns.

Freight terminal

A late truck has two prepared alternatives; one slot request is awaiting confirmation.

The handover names the consignment, pending request, last checked time, responsible operator and the action that must wait.

The next shift follows the existing confirmation request instead of issuing a duplicate or assuming the booking is complete.

Infrastructure maintenance

Discovery and preparation finished, but deployment approval did not.

The successor checks the retained plan, reviewed revision and pending approval. It resumes at the gate rather than rerunning all earlier steps.

Completed work is reused and outstanding authority remains visible.

Budget, retry & stop controls

Choose explicit attempt, time, cost or progress limits. Check them in the relevant execution path, retain the state when a limit is reached and choose a meaningful recovery or escalation. A retry policy should distinguish transient errors from failures that require a changed input or decision.

Freight terminal

A request times out repeatedly while the truck's arrival window approaches.

The coordinator checks whether the original request succeeded before retrying, limits repeated attempts and routes the unresolved case to the terminal operator.

The case retains its request identity and status instead of producing an uncontrolled stream of duplicate requests.

Research and analysis

Three successive searches repeat the same original source while the agent keeps expanding its summary.

The run records the missing evidence, stops the low-yield search path and proposes a different source or an unresolved conclusion.

Effort stays tied to information gained rather than to the number of actions performed.

Provenance & outcome evidence

Retain bounded records of relevant events and attach them to the claim they support. Keep source time, observation time and ordering distinct when they differ. A receipt for one step should not silently become evidence for the entire pipeline.

Freight terminal

A request was submitted, but the terminal has not returned its acceptance.

The record distinguishes requested, accepted and communicated. Only the acceptance record supports 'slot confirmed'.

The case summary reports the actual state and points to the specific supporting event.

Data delivery

The process returned exit status 0, but the consumer expects a particular schema and row set.

The workflow inspects the delivered output and links the schema and content check to the job revision.

Process completion and satisfaction of the delivery requirement are separately evidenced.

Brief, return & ownership contracts

Use an identified work item with a named owner, scope, inputs, expected output and return surface. Correlate the return with that item, retain unresolved findings and record the lead's disposition. A message being saved, surfaced or acknowledged describes different progress from the work being completed.

Freight terminal

The planner needs confirmation of whether pallet P12 is physically ready at the correct dock.

The request names the consignment and pallet identity, the exact check, the required returned observation and the deadline. Acknowledgment is recorded separately from the scan confirmation.

The planner can distinguish 'warehouse has the request' from 'the pallet has been verified'.

Multi-agent documentation work

Three contributors work on one guide with different evidence and file responsibilities.

Each brief names the bounded output and protected surfaces. Returns identify their source revisions and open findings; the lead records their disposition.

The guide can incorporate contributions without losing ownership or silently treating every return as accepted.

Composition / Workflows become pipelines

Inspect the connections, not just the boxes.

Each stage below is a workflow with its own input, useful output and gate. Select a stage to see where the methods enter. Then change the handoff condition to inspect how a missing or mismatched result affects the next unit of work.

Inspect this workflow's incoming handoff

A workflow can contain a review loop or a decision branch. These views describe how work is composed; they are not instructions to run every stage automatically.

Out in the world / First-party accounts

These operational questions already exist outside the lab.

The reported practices below come from the organizations' own accounts. The connections to methods in this library are an analysis of transferable problems and mechanisms; the accounts do not establish that those organizations use these named methods.

DHL Supply Chain + HappyRobot11 November 2025

Coordinating operational communication

DHL describes agents handling appointment scheduling, driver follow-up calls and high-priority warehouse coordination through phone and email interactions.

What this helps explain. This is a real setting for the kinds of source, handoff and action-boundary problems explored in the terminal scenario.

When the driver changes the arrival time, which record and which actor can actually confirm the new slot?

Read the first-party account ↗

Uber Engineering / Genie10 October 2024

Answering support questions from internal sources

Uber describes a RAG-based on-call copilot using internal documentation, source-linked answer context and a route to human support when a question remains unresolved.

What this helps explain. The useful engineering questions concern source selection, access boundaries, answer support and escalation.

Can the answer point to the relevant source, and can the user reach an on-call engineer when that evidence is insufficient?

Read the first-party account ↗

Anthropic Research13 June 2025

Research with explicit delegation and retained state

Anthropic describes a lead researcher delegating bounded work to subagents, synthesizing returns and using a citation step. Its engineering account also discusses retries, checkpoints and resumption after failures.

What this helps explain. It illustrates how a research pipeline needs coordination, bounded effort and continuity around the model calls.

What does each contributor return, and what state lets the lead continue after an interruption?

Read the first-party account ↗

Microsoft supply chain24 March 2026

Balancing competing shipment objectives

Microsoft describes CargoPilot analyzing transport modes, routes, costs, carbon impact and cycle times to produce shipment recommendations.

What this helps explain. Competing objectives create a useful setting for examining alternatives and explaining a recommendation's assumptions.

Which alternative is preferable under today's constraints, and what new information would change that assessment?

Read the first-party account ↗

Small labs / Isolate one mechanism

Where does the workflow break?

Choose a problem. Change one condition and inspect what happens. These six groups are a practical starting selection, not a universal taxonomy.

One example per group

01 / Context & orientation

Illustrative simulation · fixed rules · synthetic data

Which policy reaches the answer?

A fictional shop has two versions of its returns policy. The task concerns an online purchase, but the first document is an obsolete policy.

“How many days do I have to return an online order?”

Related: ORBIT · CARD and POINTER

Enable JavaScript to try the source-selection example. The fictional current online policy allows 30 days; the old policy allows 14.

The lesson

Routing needs scope, version and provenance. A source reference does not automatically transfer authority.

ORBIT profile ↗CARD and POINTER profile ↗

The answer is selected by a small, authored rule. No LLM runs here, and no retrieval-performance claim is being measured.

Ecosystem / Continue exploring

Parts of my harnessing ecosystem.

Public explanations provide a route into selected methods and demonstrations. Each keeps its own scope. A documented method and a working runtime are different kinds of evidence.

ORBIT

Specification profile

Orientation and source routing, with separate structural, routing and semantic relationships. The public browser lab illustrates the ideas; it is not an executable ORBIT compiler.

Explore the public profile ↗

CARD and POINTER

Semantic roles

CARD represents and orients a subject for a consumer. POINTER expresses a reference and relationship to an identifiable target. These are independent roles, not a requirement for paired files.

Explore the public explanation ↗

5PP & Dialogue Lifecycle

Work protocols

5PP structures bounded work. Dialogue Lifecycle distinguishes exploration, decisions, execution, verification and closure. The two have different responsibilities.

Explore the lifecycle profile ↗

Gate Monitor

Synthetic public demo

A public explanation of decision states using fictional events and resource values. Its private implementation is separate from the browser demonstration.

Explore the existing demonstration ↗

Worktrace

Synthetic public demo

A public explanation of action evidence and temporal boundaries. Recording an invocation does not itself prove success or the intended outcome.

Explore the existing evidence demonstration ↗

Verification & action labs

Teaching labs

The portfolio already has labs for AI evaluation, fault injection and controlled changes. They can supply deeper exercises after a short introduction here.

Browse the existing labs ↗

DIALECTIC

Interpretive loop

Develop understanding through encounters, revised interpretations and synthesis checks. The terminal case shows the specific point at which a new observation changes the plan.

Explore the existing DIALECTIC profile ↗

SORR & RigVedan

Distinct responsibilities

SORR addresses bounded identity, relationship and claim-responsibility questions. RigVedan classifies knowledge state along independent axes. They contribute different things to the same workflow.

Explore SORR ↗

Your workflow / Choose useful support

Start with the failure.
Choose the smallest useful control.

More layers are not automatically better. Identify the handoff, source, action or completion claim that needs control. Then test whether the chosen mechanism addresses it.

Explore my projects ↗About Fredrik ↗

Sources & further reading

The terminal walkthroughs, workplace examples and pipeline views are authored teaching scenarios. The four field notes link separately to first-party accounts of external practice.

The six small labs use fixed inputs in your browser. Try a different condition to inspect how the result changes. Reloading resets the examples.

Continue with the method profiles, the lab directory, or Anthropic's explanation of harnesses for long-running agents.

A capable model isa starting point.

From a method to a working pipeline.

A method becomes useful at a specific moment.

What does each part contribute?

Inspect the connections, not just the boxes.

These operational questions already exist outside the lab.

Coordinating operational communication

Answering support questions from internal sources

Research with explicit delegation and retained state

Balancing competing shipment objectives

Where does the workflow break?

Which policy reaches the answer?

Parts of my harnessing ecosystem.

ORBIT

CARD and POINTER

5PP & Dialogue Lifecycle

Gate Monitor

Worktrace

Verification & action labs

DIALECTIC

SORR & RigVedan

Start with the failure.Choose the smallest useful control.

A capable model is
a starting point.

Start with the failure.
Choose the smallest useful control.