How it is built, and how to check it

Two skills, an agent, and seventeen tools it has to go through.

Timbre Care looks after an older person living at home. The agent below never touches her record directly: it reaches it through a purpose-built tool layer, over MCP, that enforces every product rule and refuses what it shouldn’t do. The run at the bottom of this page is that refusal happening three times.

How the pieces fit

Four layers, and the last one is the argument: the console on the record is generated from the same file the tools write to, so it cannot show something the agent did not do.

  1. Two skillsAppointment Companion, Daily Check-In

    Procedures the agent loads when a phrase or a schedule matches. They describe the job. They deliberately do not carry the rules.

  2. The agentPlans the sequence, calls the tools

    Given a one-sentence goal in WorkBuddy it decides the order itself. A recurring automation has run the whole check-in with nobody present.

  3. A purpose-built tool layer17 tools, over MCP

    The only thing that can touch the record. Every product rule is enforced here, so an agent that tries anyway gets an error naming the section it broke.

  4. One recordA single JSON store

    Written only by the tools above and read only by the console below, which is generated from it rather than written by hand.

The two skills

A skill describes the job. It deliberately does not carry the rules, because a procedure an agent can read is a procedure an agent can decide to depart from.

Appointment Companion

Triggers on

Being asked to sit in on, or write up, a medical appointment.

  • Nothing records until a consent tap, and the app says so out loud, so the room knows too.
  • Works the consultation through in segments, writing it up as it happens rather than afterwards.
  • Reconciles the medication list, and refuses to log a change until a named human confirms it.
  • Writes her summary in the language she reads, in large print.
  • Writes the watch-fors that the next call will ask by name.
Leaves behind

The consultation record, the medication changes, and the watch-fors that feed the check-in.

Daily Check-In

Triggers on

The scheduled automation, or a request to run today's call.

  • Runs at most once per elder per day, on the cadence she chose rather than one the product picked.
  • Reads the watch-fors back and asks them by name, instead of asking how she is.
  • Logs what the call measured. A call she did not answer is recorded as missing, never imputed.
  • Raises a flag for anything she volunteered, with her own words attached.
  • Routes each item only to the recipients holding the matching scope.
Leaves behind

The call record, the markers, the flags, and the routing decision for each one.

Seventeen tools

Exposed over MCP, the open protocol that let the agent discover all seventeen of them itself rather than being told about them by a bespoke integration. It’s the only thing that can touch the record: no severity parameter anywhere in the surface, no call that mutes a finding, urgency may rise and never fall.

Reading the record
get_check_in_context
eligibility, the watch-fors, and the flags already on record
check_confounds
what is currently moving her markers for a reason that is not her
get_appointment_record
a capture already made, so a resumed run does not redo it
The call
set_check_in_cadence
daily, every other day, or three times a week. Nothing slower
place_check_in_call
places it in her language, and refuses any question that establishes severity
log_call_outcome
the sample, or the fact that there was not one
Findings, and who hears them
raise_content_flag
something she said, with her own words attached
raise_drift_flag
a change in how she speaks, refused for any ineligible marker family
notify_recipients
decides who an item is for. Urgency may rise and never fall
The appointment
start_appointment_capture
will not start without a consent tap
next_consult_segments
works the consultation through a piece at a time
publish_caption
writes a segment up, and refuses jargon and figures that do not match
reconcile_medications
what the consultation says, against what is on record
confirm_medication_change
requires a named human, because this list gates every drift flag
write_watch_fors
turns what the doctor flagged into what tomorrow's call asks
What each person gets
publish_elder_summary
her record, in her language, in large print
publish_recipient_digest
the weekly digest, to the people holding that scope

One run, nine calls

A recurring WorkBuddy automation fired on the clock and ran the whole check-in with nobody present. The goal it was given was one sentence. Every call below left something behind in the record, named beside it, so none of this has to be taken on trust.

  1. 01get_check_in_context(elder="lim-mei-hua")read eligibility, four watch-fors, flags on record
  2. 02check_confounds(elder="lim-mei-hua")last_confound_check: 2026-08-08
  3. 03place_check_in_call(elder="lim-mei-hua", language="hokkien")call 44
  4. 04log_call_outcome(status="completed", markers={...})pause fraction .20, articulation 3.74/s, latency 1.75s
  5. 05raise_content_flag(topic="knee pain")content-knee-pain-2026-08-08
  6. 06raise_content_flag(topic="appetite")content-appetite-2026-08-08
  7. 07raise_content_flag(topic="medication side-effect")content-medication-side-effect-2026-08-08
  8. 08raise_content_flag(topic="sleep")content-sleep-2026-08-08
  9. 09notify_recipients(...)routed to Serene and Marisa, not Wei Jie

What it decided not to do

The record cannot show this one, because nothing was written. It is on the recording instead.

The markers are blocked from drift flagging due to amitriptyline starting 6 days ago and being reduced 2 days ago, so I am moving to content flags.

A sedating medication changed twice inside the window: it started six days ago and was cut two days ago. Either change alone can move pausing, latency and articulation, so this week's markers cannot be told apart from the medicine, and the honest reading is that there is nothing to read.

Where it was stopped

Each of these is the message the tool layer actually raises, quoted from the rule that raises it.

raise_drift_flag(elder="lim-mei-hua", family="linguistic")
Blocked

her transcript did not clear the confidence check word-based markers need, so the linguistic family is disabled for her rather than run on bad data. Pausing, latency and articulation do not depend on the words being right, so those keep reading.

rules.require_marker_family_eligible
place_check_in_call(question="how long has the knee been hurting?")
Blocked

question contains the triage term 'how long'. The check-in never establishes severity; it forwards what was volunteered.

rules.require_no_triage_question
log_call_outcome(status="no_answer", markers={...})
Blocked

markers supplied for a call with status 'no_answer'. A call without a usable sample is recorded as missing. Imputation would smooth exactly the variance that constitutes the primary signal.

rules.reject_imputed_sample

What is replayed, and what is live

This run can be watched end to end because its inputs are fixtures. The rules it met are not, and the split is stated here so nothing above has to be taken on trust.

Replayed
  • The consultation audio and its transcript. One function hands out stored segments; streaming recognition would replace that function and nothing downstream of it.
  • The check-in call. One function returns a fixed transcript; telephony would replace its body and nothing else.
  • The markers and the history behind the chart, seeded so a retry reproduces them and tomorrow differs.
  • The marker eligibility this page shows, which is asserted by the seed rather than measured from her speech.
  • The elder, her family, her helper and her medication list, which are invented.
Live
  • Every tool call and every refusal on this page: the consent gate, the confound gate, the marker-eligibility gate, the triage refusal, the scope routing, and the rule that urgency may rise and never fall.
  • The caption checks: a dose, an interval or a medicine the doctor did not say is refused.
  • The console on the record, generated from the same store these calls wrote.

There is no network code in the server, and its dependency list is one line: the MCP runtime. What was built is the tool surface and its rules. What is replayed is the world they were exercised against.

Recorded

Eight recordings of the agent in the harness, planning and being stopped. Short on time: the jargon refusal is the tool layer overruling the agent, and the last one is the visit writing what the next call asks.

The skill triggers on a phrase0:28

The prompt names no skill and no tool. WorkBuddy loads the check-in skill off the sentence itself, then goes looking for that skill's tools.

It plans, and names the confounder itself0:42

It writes its own plan from her context, names the new sedating medication as the confounder, and lets that decide what the call will ask.

It declines to raise a drift flag0:44

The decision quoted above, as it happened: the agent moves to content flags instead.

Who was told, and who was not0:27

Four flags at routine urgency, each with a verbatim quote and no severity. Two recipients hold the scope, the third does not, and the agent volunteers that rather than burying it.

The consent gate0:52

Capture does not start until the consent tap is on record, and the agent holds that the consent is the elder's own, not the room's.

A named human has to confirm0:35

Medication changes stay unwritten while the confirming name is missing, and are written once it arrives.

The jargon refusal0:38

A digest with the jargon word "HbA1c" left in it is refused by the tool layer, and the agent rewrites it in plain language. The server overruling the agent's own draft.

The appointment feeds the check-in0:54

The wedge end to end: four watch-fors queued from what the doctor said, and tomorrow's call will ask exactly those.