alexcinovoj.devby TechTide AI
Back to proof
Track · AI production systems

What has to be true before an agent runs in production.

Ten implementation patterns, each with the failure modes it exists to prevent and the checks that prove it holds. Every pattern here is applied in at least one of the project pages below.

Patterns

Acceptance tests for agent work

Define done before the agent starts.

How to write acceptance tests that decide whether an agent run shipped something correct, rather than something plausible.

Evals that gate releases

A golden suite in CI, not a spreadsheet.

Running an adversarial eval harness against a golden task suite as a release gate, with published pass thresholds.

MCP permission scoping

Deny by default, declare every tool.

Scoping a Model Context Protocol server so a caller reaches only what it declared, with read-only hints and no ambient credentials.

Observability for agent runs

One trace ID from request to receipt.

Instrumenting agent runs so a failure can be reconstructed without asking the model what happened.

Cost controls that actually stop spend

Ceilings enforced before the call, not after the invoice.

Per-run, per-agent, and per-caller budget ceilings enforced in the execution path.

Approval gates that suspend the run

Nothing downstream executes before a human approves.

Human-in-the-loop gates implemented as run suspension rather than as a notification.

Trace replay and diffing

Replay the run, diff the divergence.

Replaying a recorded run deterministically and diffing two runs to find the first point of divergence.

Recovery and durable checkpoints

A crash resumes; it does not restart.

Checkpointing long agent runs so an interruption costs one step rather than the whole run.

Rollback for agent-made changes

Every write has a documented reverse.

Making agent-authored changes reversible: change sets, inverse operations, and a rehearsed rollback path.

Ownership handoff

The team operates it after the engagement ends.

What has to exist before an AI system can be handed to an internal team: runbooks, dashboards, named owner, and a decommission path.

Where these are implemented

ClawKeeper

Governed Claude agent control plane. A deterministic policy engine runs before every LLM call, autonomous spend defaults to zero, and the audit log is enforced by the database.

TechTide Harness Kit

Trust contracts for agent skills expressed as JSON Schema: risk tiers, tool permissions, deny-by-default egress, approval gates, and prompt-injection fixtures behind a wall of CI validation gates.

CipherClaw

Deterministic multi-agent debugger with zero LLM calls in the analysis path: causal graphs, weighted root-cause scoring, and failure prediction.

Swarm 357

357 Claude agent roles across six business layers, with portable memory, enforced cost caps, a bash policy gate, and durable checkpoints.

FintheFinder

Deep-research agent whose runs suspend until a human approves, with a deterministic citation audit that keeps AI out of the verification path.

TechTideAI2

Company-scale agent platform on a three-plane control/data/evidence architecture, with state-machine transitions, hard execution timeouts, and an adversarial eval harness over a 33-task golden suite.

BRI

Video intelligence service with a Streamlit operator surface, a FastAPI MCP server, SQLite durability, and multimodal tooling.

AgentRewind

Trace forensics build: zero-trust import parser and a diff engine for replaying agent runs. Private build with no public artifacts.

Molten

Governed agent runtime: kill switch, circuit breakers, four-tier sandbox, dead-man's switch. Private build with no public artifacts.

Every pattern in this track

Engagement scope, process, and pricing live on TechTide AI. This domain only carries the technical proof.

AI Production Systems at TechTide AI