Detect, history
A habit in the history
Every Monday a community organiser asks for a summary of the week. Nobody designed an agent for it.
Research / Context track / Context Lifecycle Engine
Agents that emerge from how you work, and lose their standing when proof expires
The Context Lifecycle Engine watches a user’s prompt history, detects intents that deserve their own agent, builds each candidate into a content-addressed image, validates it on the user’s own past, and moves it up or down a lifecycle as evidence accumulates or expires. Follow one Monday habit through the whole system.
- Two hashes
A candidate’s source hash is never its built image hash. Lifecycle tags attach to images only.
- Goodhart boundary
A container has no read path to its own metrics. They are written one way, to a system-owned volume.
- Failed builds consume nothing
A failed resolve, replay or assemble leaves the store byte-identical.
- One JSON line per operation
With a mandatory evidence field on any upward move. No log, no merge.
- Replay proves the trigger, never the answer
Replay outputs are pre-evidence and can never flow into a promotion.
- Proof expires
Images freeze a model fingerprint. When the served model drifts, the agent is demoted.
The flow
A Monday habit becomes an agent
Detect, episodes
Cut into episodes
Silence and explicit markers decide where one attempt ends and the next begins.
Detect, clusters
Same intent, different words
Each episode’s opener is embedded and clustered incrementally. Thresholds are relative to this user’s own baseline, never absolute.
Detect, signal
Recurrence, not failure
Retried within minutes, the same intent means the topology fails this user: a repair. Spaced at a stable period, it is a routine: an automation.
Detect, stability
A check that admits what it cannot see
An unstable cluster yields no candidate. When the check cannot run in this vector space it says
unavailable, and the candidate is born carrying that gap for the human to see.Build, candidate
A YAML of references
No new logic: every tool, memory scope, prompt fragment and model is a hash into a library of already-evaluated components, plus a provenance block that answers “why do you exist?”.
Build, gate
A human approves the birth
Synthesis is optimistic, activation is gated. A rejected candidate stays in the store as a documented experiment.
Build, resolve
Every reference must exist
A missing hash fails the build in milliseconds and consumes nothing. Declared tools must be tools, and a trigger’s required capability must be mounted.
Build, replay
Validated on the user’s own past
Replay measures how many of the cluster’s episodes the trigger would have caught, how much other traffic it would have stolen, and what the habit cost. It proves the trigger, never the answer.
Build, assemble
Compiled and fingerprinted
The prompt is compiled in declared order and the served model’s fingerprint is frozen into the artifact.
Run, image
Two objects, two hashes
The image is hashed as a whole. Its trigger, the intent centroid and its period, is baked in: if the habit drifts, a new version is built.
Run, containers
One image, many lives
Each workspace runs its own container of the same image.
Run, mounts
The world is in the mounts
Memory scopes read-only or read-write, MCP connections. The agent is cattle, and its volumes are the pets.
Run, metrics
Scores it cannot read
Metrics are collected per container and written one way. An agent that could observe its promotion criteria could game them.
Lifecycle, tags
Promoted on lived evidence
A human moves it to trial, then to ephemeral with real occurrences: cost ratio 0.6 over four Mondays, all closed as successes.
Lifecycle, shadow
The engine judges in shadow
In v1 the state machine runs beside the human and logs what it would have done. The disagreements are the calibration set for turning it on.
Governance, oplog
Every move is a line
One JSON line per operation. An upward move without an evidence field is refused.
Governance, topology
The learned topology is a file
Every change to
topology.yamlis a commit.cle logshows who approved what, on which numbers, andcle diffshows what usage grew.Lifecycle, re-validator
The model drifts
The re-validator replays the image’s probe set against the served model. The fingerprint no longer matches.
Lifecycle, tags
Proof expires
The agent drops back to trial and the oplog records
revalidation_failed. It has to earn its standing again.
Measured
What the corpus said
- 32 / 48
intents grouped as Stack Overflow moderators did, against about 2% at random.
- 0.08%
of real WildChat users have enough repeated history to detect anything. The engine is for intensive users, by design.
- ~10
occurrences of one intent for reliable recovery. Episode count predicts nothing.