It knows when it’s stuck. And what to do about it.
A deterministic layer beside every agent. It reads every step, keeps five feelings as working state, and when one crosses a line, changes what the agent does next.
The emotion engine running one Bean Street turn, on a loop. Events land in a feed on the left (the brief read, an export that fails twice the same way, a discount Anika asks for that the brief rules out, Priya declining it, a deadline a day closer, the task budget 80% used, twelve checks passed), travel along a beam into the core, where twelve readings move on a ring around the Foxora mark and five feelings keep their levels, and when a level is entered one short host note goes out to the crew (the engine is shown in act mode): change approach, verify first, slow down and check in, finish the core. A rail along the bottom names the phase: sense, appraise, feel, act, show.
- Status
- In Studio V6.0 · observe mode
- Feelings
- Five, plus calm
- Readings
- Twelve, from seven private signals
- Model calls to feel
- None
Work has a state. Now it has a name.
Each feeling is a reading of the work, not a judgment. Each one changes what the agent does next.
Calm
“All good so far.”
- Felt when
- the other four at 0
- So it
- Works through the plan, one step at a time.
Frustration
“I’m stuck.”
- Felt when
- same failure, twice
- So it
- Stops retrying and tries a different approach.
Pressure
“Time’s running out.”
- Felt when
- budget or context low
- So it
- Finishes the core and reports what’s left, instead of cutting corners.
Doubt
“Is this right?”
- Felt when
- results unclear or disagree
- So it
- Verifies before it claims anything.
Caution
“I’ll re-check my work.”
- Felt when
- you said no, or stop
- So it
- Doesn’t retry what you declined, and re-checks its work before going on.
Confidence
“This is working.”
- Felt when
- checks fresh and passed
- So it
- Keeps going, and never skips the final check.
Behind the five · twelve readings
- calm
- focus
- curiosity
- confidence
- relief
- frustration
- pressure
- anxiety
- doubt
- caution
- fatigue
- resilience
Twelve readings behind the five: pressure carries half of anxiety, doubt carries anxiety and fatigue, confidence carries relief and resilience.
One bounded note. Never a new power.
When a level is entered, the agent gets one short host note at its next model boundary. Four per lead, two per specialist, six per turn.
sheets.export · failed twice · same signature
Frustration· L2
Foxora host note, from this run’s events, not from the person: The same sheets.export operation failed more than once. Use the observed failure to choose a different approach before retrying.
1 of 2 for this specialist · 3 steps since the last · no tool closed, no approval added
result · operation_recovered
Two more actions are named in the protocol and not sent in this version: pause and report, hand back.
A feeling never overrides your rules.
A note can change how an agent works. It cannot change what an agent may do.
- 01
Never closes a tool
A note can say change approach. It cannot take a tool away.
- 02
Never adds an approval
The six actions that ask in Auto still ask. A feeling never adds a seventh.
- 03
Never changes a permission or mode
Ask, Accept, Auto and Full Access are yours. The engine reads them, it never writes them.
- 04
No model call to feel
Appraisal is arithmetic over the run’s events: no prompt, no extra request, no latency.
- 05
Private stays private
Seven signals never leave the engine. Details shows levels and readings, not your words.
- 06
Never carries a fact, only a feeling
A lead keeps its feelings across the messages of a task. Evidence, checks, approvals and note budgets start fresh every turn, and nothing becomes a judgment about you.
Observe mode by default. Act mode sends the notes. Off turns it all off. One launch setting, read once per turn.
It never claims more than the run can show.
Before a lead’s reply goes out, an audit checks every claim against what actually ran. One correction, at most one retry per turn, tone untouched.
Audit · from the run’s facts
- checks · 11 passed · 1 skippedfresh
- push · never ran · needs your yes
- send · not attempted
retry · 1 of 1 this turn
Tone untouched. Only the claims moved.
One line in the thread. The rest in Details.
Observe mode reads every run and shows its working state. Act mode sends the notes, and becomes the default only once held-out tasks show fewer wasted retries at no more cost.
- Today
- Preview Observe mode, in every local task
- Act mode
FOXORA_EMOTION=act- Resets
- Facts, checks and note budgets, every turn. A lead’s feelings carry across the messages of a task.
- Not covered
- Shared, channel, voice and cloud work
A Foxora Studio window on the Bean Street proposal with its Details panel open. Researcher’s export of the order sheet fails twice the same way, so the engine’s frustration reading enters level 2, a change-approach note is sent (the window is shown in act mode), and Researcher exports as CSV instead; the readings then ease back toward focus while Writer drafts the proposal. The panel shows the primary label, the five levels as bars, the twelve readings on a radar and a short history of what was observed and sent.
A good teammate notices. Now your agents do too.
Five readings of the work. Each one a reason to change course, none of them a new power.
Part of Foxora Studio V6.0. See it on the homepage
Free plan · 200 credits a month · No card required