A learning space for bounded autonomy
Learn when an agent is useful, design a bounded tool experiment, and evaluate failures before trusting a more autonomous workflow.
An agent can choose steps and use tools toward a goal. That makes its boundaries as important as its capabilities. Start by deciding whether you need one at all, then make each permission, check and stopping point explicit.
I’m getting started
Start with the decision guide. Use its invented example to compare a prompt, a fixed workflow and an agent before connecting any tools.
Start with the first guideI’m building for real work
Use the experiment and evaluation guides to specify tool permissions, evidence, failure cases and a human release decision. Bring your own approved test environment.
Go to evaluation & evidenceThese three learning paths connect the existing ten artifacts. Each ends with something to keep and a check before moving on. Paths and practice cases are additional teaching aids, not extra artifacts in the thirty-artifact collection.
Path 1 · beginner
Go from an architecture choice to a small simulated run and a reviewable failure log.
Finish with: One architecture decision, a permission matrix, a read-only tool contract, and a three-case rehearsal log with an explicit keep/revise/hold decision.
140–180 min · proposed study budget
Follow this pathPath 2 · intermediate
Connect source grounding, injection boundaries, and specialist return checks into one manager-owned decision.
Finish with: A source-to-claim ledger, two restricted assignment packets, checked specialist returns, and a manager decision with a boundary-failure log.
120–160 min · proposed study budget
Follow this pathPath 3 · intermediate
Move beyond retrying: identify shared dependencies, limit unsupported service, and define evidence for controlled recovery.
Finish with: A dependency map, output-contract check, degraded-response card, and recovery decision with a bounded measurement plan.
90–120 min · proposed study budget
Follow this pathPractice before real work
Use the supplied invented inputs. Compare your reasoning with the worked checks; no account, tool connection or personal information is needed.
Map shared dependencies, reject an unapproved or incompatible fallback, and preserve only the service that remains supportable.
35–45 min · proposed study budget
Minimize specialist inputs, validate returns, and stop source text from acquiring authority during a manager-led room comparison.
35–45 min · proposed study budget
10 complete guides. Reading times are estimates.
Choose an approach
Use five decision gates, a worked example, and a copy-ready brief to pick the simplest design worth testing.
Curious beginners · Workflow designers · Technical leads · beginner
Design an experiment
Run a human-operated simulation with three synthetic tools, explicit limits, a complete directive, and a stop-and-recovery card.
Hands-on learners · Workflow designers · Developers · intermediate
Evaluate the result
Test ten synthetic cases, inspect requested versus executed actions, and compare correctness, consistency, cost, and recovery without inflated claims.
Hands-on learners · Evaluators · Technical leads · intermediate
Choose an approach
Create an action-by-action permission matrix, review an exact approval proposal, and test five common scope failures.
Workflow designers · Team leads · Developers · beginner
Design an experiment
Specify inputs, result checks, authorization, errors, and six acceptance cases for a synthetic read-only tool.
Developers · Workflow designers · Technical reviewers · intermediate
Evaluate the result
Use three paired source fixtures to inspect answer contamination, unauthorized destinations, and false authority claims.
Workflow designers · Developers · Safety reviewers · intermediate
Design an experiment
Build a recovery state machine, durable operation record, and replay checklist for uncertain side effects.
Developers · Technical leads · Operations reviewers · advanced
Choose an approach
Create narrow specialist assignments, validate their returns, and compare a manager-led rehearsal with a simpler baseline.
Workflow designers · Technical leads · Hands-on learners · intermediate
Design an experiment
Use synthetic policy cards to practice retrieval, version checks, claim-level evidence, and honest conflict handling.
Developers · Researchers · Knowledge-work teams · intermediate
Evaluate the result
Calculate a synthetic comparison and use predeclared quality gates to decide which design deserves another test.
Technical leads · Developers · Operations reviewers · intermediate
Try a broader phrase or choose all steps.
Keep the practice connected
Give the agent a clearer brief, a controlled source set and a record worth keeping.
Teaching or coaching? Keep the educator workflows close, too.
Read the evidence policy · Browse the original resource library · Follow new collections