How agents are born, promoted and retired

Part 12025-08-11theory

Context as a Service

A second question: maintaining one agent is feasible, maintaining a hundred is not. The answer is to stop writing agents as monoliths.

• The Context Block: declarative and immutably versioned, holding a model, tools, a memory scope, logic, and its own evals.
• Application, environments, blocks, with consistency tests between the blocks of an environment.
• Two levels of versions, block semver and an application snapshot. Promotion from dev to stage to prod, with atomic rollback.
• Definition on the platform, invocation in code: a new prompt ships without a redeploy.
• Point the code at a floating latest, or pin an immutable version.
• Any model provider: the block holds a pointer, not a dependency.

Article (EN) Article (FR)

Parts 2–52026-01 → 04from the memory track

Borrowed from the memory track

Part 7 lists what each earlier article contributed. Three pieces cross over from memory to lifecycle.

• From User-Centric Context Engineering, the trust model: write optimistically, let a human correct afterwards.
• From UX-Driven Agent Memory, a law: typed blocks earn permanence through use, from temp to project to fundamental.
• From Agent OS, the Force Override: a human can always interrupt.
• From the APU, a measured baseline: 78% of queries served by L1.

The memory track

Part 6forthcomingdraft

Context Switch Engine

The link between switching contexts and creating them. The article is still being written, but parts 7 and 8 already build on it.

• A Merkle DAG context store: every context state is a commit of content-addressed blocks, and a switch loads only the diff.
• A rupture detector that cuts a conversation into episodes, later reused by the CLE.

Described from parts 7 and 8 until it is published.

Part 72026-07-16theory

Context Lifecycle Engine

Every multi-agent system fixes its topology at design time, for every user, forever. The CLE lets it grow from each user’s usage instead.

• The developer assembling blocks. A synthesiser composes candidates from a library of hashed, already-evaluated components.
• Environments as the promotion path. Seven states replace them, from pattern to pinned, then deprecated, archived, and resurrected.
• A signal that needs no prediction: iterations before a result. Reformulation triggers a repair, recurrence an automation.
• Evidence required on every upward transition, and an abandonment guard: a user who gives up is not a success.
• The Force Override becomes a maternity ward: a human approves every birth.
• Auto-approving candidates built only from pinned components: possible later, out of scope for v1.

Article

Part 82026-07-17theory + code

Build, Ship, Run

Reread the next day, part 7’s YAML was a Dockerfile. Docker is not the model: it shares the same ancestors, and gives a vocabulary everyone already holds.

• A build step: resolve every reference, run the components’ evals, assemble and hash. Definition errors fail before any user sees them.
• The YAML and the image become two objects with two hashes. Lifecycle tags attach to images.
• Image and containers: metrics per container, decisions per image. Volumes hold the state, and the agent cannot read its own scores.
• topology.yaml, written by the CLE and audited by humans. Governance is git log topology.yaml.
• A trigger adjustable at runtime. Rejected: the trigger is the agent’s identity.
• Namespaces, cgroups, Docker networking. Not taken: the analogy is structural, not a security claim.

Architectures are mines, not templates.

Article How the CLE works, step by step EZFRICA/context-lifecycle-engine code the next day

Measurement2026-09-06measurement

The cost of trusting AUC

The measurements part 8 promised, on pairs humans had already judged: Stack Overflow questions closed as duplicates by moderators.

• A norm guard at the boundary that refuses a vector from a foreign embedding space instead of normalising it.
• A dedicated vector database. Not justified: exhaustive search stays under 3 s at a million vectors.
• BigQuery for measurement only. The engine and its tests never touch it.

Benchmark your operating point, not your model.

Article