Outcome Completion Agent

Outcome Completion Agent

Use this skill when the user states an outcome rather than a call, the route to that outcome runs through more than one organisation, and the next organisation is usually named on a call rather than known in advance.

"Get a replacement pallet for the crushed delivery on order BK-7741, before Friday, under $500."

The user has one phone number: the supplier who sent the broken pallet. The supplier is out of stock and names a distributor. The distributor quotes over the limit and names its own depot. The depot can do it. Three organisations, two of which did not exist as far as the agent was concerned when the run began.

outcome-completion-agent is a calling-pattern skill. It adds no backend, no queue and no daemon. It turns one authorized "make this happen" request into a bounded loop of one-off CALL-E calls whose order is decided by what earlier calls established.

When To Use

• a problem whose owner is unknown: "find out who can actually fix this and get it fixed"
• a chain of custody: supplier to distributor to depot, carrier to facility to counter, insurer to provider to adjuster
• a goal with a hard limit attached — a price ceiling, a date, a required reference — where a cheerful "yes" that breaks the limit is a failed outcome
• any request where the honest answer to "who should we call?" is "ask the first person"

When Not To Use

• one call, one structured answer. Use a single-call skill; this loop is overhead.
• a known, priority-ordered candidate list for one opening. Read skills/priority-call-waterfall/SKILL.md instead — one opening, call in order, stop at the first yes.
• comparison shopping over an enumerable set of same-service vendors (every barber within 4km). If the candidate list is a database or map query, the frontier does not need to grow and this skill's machinery buys nothing.
• anything email, a web form, or an API can do. A phone call is a real-world side effect and the most expensive way to ask a question.
• goals with no completion test. If nobody can say what "done" looks like, the loop cannot stop, and a loop that cannot stop must not be given a phone.

Core Workflow

1. Confirm the user wants this pursued by phone, now, and that they accept calls may go to organisations they did not name.

2. Collect the outcome fields (see Required Fields). Ask for anything missing. Never infer a phone number, a price ceiling, or a deadline.

3. Show a masked preview: the goal, the constraints as the agent understood them, the starting phone book, and the call budget. Get explicit confirmation before call one. Always show the parsed constraints back. A price ceiling that failed to parse is a ceiling that will never block anything, and the user is the only one who can catch it.

4. Loop, one call at a time:

   plan -> (approval, if it commits) -> one call -> evidence -> constraints -> plan
   

   • plan — pick the next party from the frontier: a party never called, else a party that did not answer and is under its retry cap. Prefer a party named by the call that just happened over one the user listed but nobody has reached.
   • call — exactly one CALL-E call, with a structured result schema. Wait for a terminal status. Read references/evidence-schema.md for the schema every call fills.
   • evidence — record what was established: verdict, facts, blockers, referrals, and any concrete offer. Add referrals with a valid E.164 number to the frontier. Drop referrals without one.
   • constraints — evaluate any offer against the user's limits. Read references/constraint-evaluation.md. A call that completed happily and breaks a hard limit is a failed outcome: keep working.

5. Stop conditions, checked before every dial:

   • an acceptable offer has been confirmed by the other party → resolved
   • the call budget or the per-party cap is reached → stop and report
   • every lead is exhausted with nothing acceptable → stop and report

6. Exactly one call in the run may commit the user, and only after they approve it. Read references/approval-gate.md.

7. Report using the Output Format below, whether or not the goal was reached.

Required Fields

• goal — one sentence, imperative, with the constraints taken out of it
• constraints[] — each with kind, description, value, and hard (budget · deadline · required_fact · preference · forbidden). hard: true disqualifies an offer. hard: false only ranks it.
• organizations[] — the starting phone book: name, phone (E.164), role. One entry is enough; that is the point of the skill.
• budget — max_calls for the whole outcome and max_calls_per_org. Both are required. A goal-seeking agent with an unbounded phone budget is not a product, and CALL-E accounts are metered.

Phone numbers must be E.164, and must be masked in every summary, preview and report.

Following A Referral

This is the part that distinguishes this pattern, so it has its own rules.

• A referral counts only when the person on the call named a number. "Try their head office" without a number is a fact, not a lead.
• Validate it as E.164 before adding it. A number that cannot be dialled is not a lead, and keeping it puts an entry in front of the planner that can never be actioned.
• De-duplicate on the number. Two calls naming the same depot must not produce two entries, or the depot gets called twice.
• Carry the provenance: which call produced this lead, and the reason given. The report has to be able to say "the agent found this one", and the next call should be able to open with "Halden Packaging suggested I call you".
• A referral is a lead, not an instruction. It still costs a call from the budget, it still goes through the same constraint checks, and it never bypasses the approval gate.

Budgets

Check the budget immediately before dialling, never when planning. An approval can sit overnight; the credit it would spend may be gone by the time it comes back.

Sensible defaults for a metered account: 6 calls per outcome, 2 per organisation. The per-organisation cap is what stops a no-answer loop from consuming the whole budget on one unanswered line.

Never widen a budget mid-run to finish a goal. Stop, report, and let the user decide whether it is worth more calls.

Output Format

Report every run, resolved or not:

• the goal, restated, and the constraints as they were evaluated
• every organisation touched, in order: masked number, calls used, final verdict, whether the agent found it or the user supplied it
• every offer received, with a per-constraint verdict and, for the ones turned down, which constraint they broke
• the outcome: what was secured, from whom, at what price, by when, and any reference number — or why the run stopped
• calls used against the budget

Never report an outcome as resolved unless the other party confirmed it on a call the user approved. An offer is not a resolution.

Safety Rules

Read references/safety.md for the full contract. Always:

• Disclose that the caller is an AI assistant, at the start of every call, before asking anything. If the person asks for a human or asks you to stop: apologise, end the call, record it, and do not call that number again in this run.
• Gathering calls have no authority. Their script must say so explicitly.
• One call per run may commit the user, only with prior approval, and only to the exact terms approved.
• Mask every phone number in output. Never read a credential aloud.
• Medical, legal, financial and emergency matters are logistics only: ask about times, availability and reference numbers, and give no advice.
• This skill creates no recurring schedule. Every call is one-shot. To stop a run, stop approving; nothing is committed without an approval, and nothing is queued for later.

Worked Examples

Read references/examples.md for a full run, a run that stops on budget, and the two failure modes that most often get this pattern wrong.

Validating A Result

scripts/check_evidence_schema.py checks one structured_result against the schema and reports what a planner would actually be able to use from it — no network, no credentials, no calls:

python3 scripts/check_evidence_schema.py path/to/result.json

It reports errors (the result is malformed) separately from notes (the result parses, but something in it will be discarded). The notes matter more than they look: a referral whose number is unusable is the difference between an agent that finds the depot and one that stops at the supplier, and nothing else in the run will tell you it was dropped.

Check the checker with python3 scripts/test_check_evidence_schema.py.