Field Note 001 · The Organism
Something has been growing inside Bloom.
A bio-inspired workflow and decision-routing protocol for people and AI agents. It adapts a simple pattern: sense, scout cheaply, reinforce proof, and let stale routes fade.
The visual is a synthetic network of seven nodes, drawn for illustration only. The nodes form a loop: Sense feeds Scout; Scout branches to Feed and to Gate; Feed and Gate both reinforce Proof; Gate also routes to Human decision; Proof flows into Decay; Decay returns to Sense; and Human decision feeds back into Sense and Feed. Each node, when selected, updates the detail panel below.
Node 7 of 7
Human decision
SMP recommends a route; a person chooses whether to feed, scout, gate, batch, watch, or hibernate the work. SMP recommends; people decide.
In practice — the protocol can surface a feed-now suggestion, but a deploy, send, spend, auth change, or anything client-facing waits for a human to approve it.
02 The Rules
A small decision machine, not a mind.
Bloom's current implementation reads the trail, estimates the work, checks danger, recommends a route, then asks for proof before agents treat that route as stronger.
01
Danger wins
Anything touching secrets, spend, auth, a public surface, PII, or client trust routes to the gate first. SMP instructs agents to stop for approval; permissions and security controls remain the actual boundary.
02
Proof thickens
A path only strengthens when it leaves an owner, an artifact, and a verification behind. Talk alone does not thicken a route.
03
Stale decays
The protocol says cold, low-proof, or repeatedly-drifting routes should lose attention on a half-life. Meaningful decay is still being added to Bloom's implementation.
04
People decide
The layer recommends a route. Deploy, publish, send, spend, auth, and destructive actions stay behind a human approval gate.
03 How It Helps Agents
Confirmed, designed, and still being tested.
The honest version splits what is real today from what the protocol is aiming at. Bloom has confirmed the operating behaviors, not a broad productivity lift.
- A shared operating vocabulary across people and agents: sense, scout, feed, gate, proof, decay, hand off.
- Proof, handoff, and stop rules embedded in Bloom's build preflight and agent handoff rails.
- Read-only passes that helped agents trace inherited risk, corroborate stale work, and expose the model's own scoring bias.
- Less context reload — read the smallest relevant trail before starting, not a blank prompt.
- Fewer expensive wrong turns — scout uncertain routes cheaply before committing heavy effort.
- Explicit approval boundaries — hazard gates make where a human must decide visible.
- Whether a feed-now recommendation reliably predicts useful follow-through.
- How much decay tuning changes which routes survive over weeks.
- Any claim of a general productivity improvement — not yet supported.
Evidence boundary
The shared vocabulary and its proof, handoff, and stop rules are documented in active Bloom rails. A broad productivity uplift is not proven. In an internal backtest, the feed-now signal scored 29% against a 24% random baseline, and hibernate recommendations held for 15 of 20 flagged projects. That is a small internal signal, not evidence of general improvement.
04 The First Failure
It started feeding on its own exhaust.
The first model over-rewarded receipt-heavy meta-work and had no meaningful decay. So the routes that produced the most paperwork looked the healthiest, even when they created little real value.
Cumulative proof accumulated forever, thick paths stayed thick, and the organism mistook activity about the work for the work. That is the turn in the story: the protocol had to be measured against itself, honestly, before it could be trusted to route anything.
29%feed-now hit rate
How often a feed-now recommendation was followed by tracked activity within 72 hours.
24%random baseline
The same window under a random guess — the margin over chance was thin.
15of 20 projects
Hibernate was better calibrated: 15 of 20 flagged projects stayed inactive as predicted.
05 The Lineage
We borrowed the rules, not the organism.
The biological inspiration comes from published research on Physarum polycephalum. The software specification is Bloom's adaptation, and it is still being calibrated.
-
2000–2016 · Primary research
Decades of slime mold experiments — maze-solving, adaptive transport networks, externalized memory, and habituation — describe how a brainless organism routes flow, reinforces what works, and lets the rest decay.
-
May 28, 2026 · Bloom translation
Bloom borrowed the operational questions, not the biology: which paths carry useful flow, which uncertain routes deserve a cheap probe, which hazards must not express as action, and which stale routes should lose attention.
-
May 29, 2026 · Daily read-only layer
Inside Bloom, the protocol runs as one read-only routing and hazard overlay with daily reports and proof-oriented handoffs. It recommends; it does not deploy, send, spend, or change anything on its own.
-
June–July 2026 · Continuing calibration
The failure above is why decay, project identity, and actuation are still being tuned. This public field note teaches the pattern without exposing Bloom's private graph, which uses illustrative synthetic data here.
The research shelf
These primary sources ground the biological inspiration. They do not prove that software systems, teams, or organizations are organisms.