Harness engineering

The five load-bearing layers around a model.

Layer 1 — small, safe, reversible actions; exact edits over fragile diffs.

Layer 2 — tests, coverage, SAST, secret scan, AI review, CI must all agree.

Layer 3 — no blind rewrites, no secrets, no new external calls, revert on fail.

Layer 4 — measure the loop's own health honestly.

reflect -> propose -> gate -> ship -> verify -> auto-revert -> learn.

A change that fails CI after shipping is reverted automatically.

Isolated, permission-scoped execution environments (containers, microVMs like Firecracker/gVisor) so an agent's tool…

Defending agents against prompt injection, jailbreaks, tool misuse, and data exfiltration.

Scoped credentials, least-privilege permissions, and auditable identity for autonomous agents and each of their tool…

Attacks against AI systems — prompt injection, jailbreaks, data poisoning, model extraction, and adversarial examples.

Deliberately encoding rules, voice, and constraints into the instruction text fed to a model to steer its behavior…

A guardrail that blocks the self-improvement and idea loops from shipping changes or ontology concepts that duplicate…

A standard (SPIFFE) and runtime (SPIRE) that issue cryptographic workload identities (SVIDs) via node/workload…

The loop is driven by time-triggered cron jobs (e.g. ~15-minute Spark telemetry refreshes and the self-improve loop) so…

The guardrail principle of limiting how much sensitive data is exposed and for how long, re-hiding it as soon as it is…

An observability technique that reconstructs and visually plays back a coding agent's actions step-by-step over a…

A technique that records an agent or program's full execution so a developer can step both backward and forward through…

The deliberately designed tool/command surface an LLM agent acts against, optimized for model legibility and…

Böckeler's harness framing: GUIDES steer the agent before it acts (feedforward); SENSORS catch mistakes after…

How amenable a codebase or environment is to agent control mechanisms — raised by 'ambient affordances': structure that…

Deriving every displayed value from exactly one canonical origin so redundant copies cannot drift out of sync.

The hidden, fixed token cost an agent harness adds to every model call through system prompts, tool schemas, and…

A lifecycle-hook plugin (github.com/DietrichGebert/ponytail) that makes a coding agent think like 'the laziest senior…

A pre-generation guardrail pattern: force an agent down an ordered ladder of cheaper options before it writes new code…

The recurring engineering practice of porting an agent onto a newer base model and re-benchmarking its latency, cost,…

A checked-in natural-language file (e.g. CLAUDE.md or AGENTS.md) that encodes a repo's conventions, guardrails, and…

Treating an agent run's context and output tokens as a finite, explicitly allocated resource, spending it on useful…

Persisting an agent's durable knowledge and state as plain files in the repository or filesystem rather than in the…

A harness layer that decouples agent logic from any single model provider so models can be swapped without rewriting…

Automatically stripping secrets, tokens, and sensitive values out of execution logs and telemetry before they are…

A TLS-terminating man-in-the-middle proxy that decrypts and logs an application's outbound traffic to reveal exactly…

A cooperative advisory lock (e.g. /tmp/darkfactory-ontology.lock) that autonomous loops must acquire before rewriting…

The design paradigm of building small, single-purpose, composable tools that each do one thing well and pipe together,…