How I test agent work and investigate failures.
I document testing, permissions, approvals, and recovery methods. Each note describes checks to try and ways the approach can fail. Related project pages link to source code where it is public.
Patterns
Acceptance tests for agent work →
Executable checks test the task and detect regressions.
I explain how executable checks test an agent’s requested task and detect regressions.
Evaluation tests for releases →
A fixed evaluation suite checks release requirements.
I explain how to run a fixed evaluation suite and enforce a release threshold.
MCP tool permissions →
The server checks which tools a caller may use.
I explain how an MCP server checks tool permissions for each caller.
Recording agent runs →
A shared trace ID connects the records of a run.
I explain how shared trace IDs connect tool calls, retries, and results for debugging.
Spending limits for agent runs →
Budget checks run before provider requests.
I explain how to check spending limits before issuing provider requests.
Pausing for human approval →
The run waits for approval before the action executes.
I explain how to pause an agent run until a person approves the proposed action.
Replaying and comparing agent traces →
Step-by-step comparison finds where recorded runs differ.
I explain how to compare recorded runs and locate the first differing step.
Recovering runs from checkpoints →
Stored checkpoints support recovery after interruption.
I explain how stored checkpoints and repeatable steps support recovery after interruption.
Reversing agent-made changes →
Each change needs a documented reversal procedure.
I explain how to record changes and rehearse their reversal procedures.
Handing a system over to a team →
A named owner practices running, debugging, and stopping the system.
I explain how a named owner practices running, debugging, and stopping a system.
Related engineering projects
ClawKeeper →
ClawKeeper checks proposed Claude agent actions against explicit rules before a model call. Autonomous spend defaults to zero, and the database enforces an append-only audit log.
TechTide Harness Kit →
TechTide Harness Kit uses JSON Schema to declare skill permissions, risk tiers, network access, and approval requirements. Automated checks validate declarations and run prompt-injection fixtures.
CipherClaw →
CipherClaw analyzes recorded agent activity without model calls. It builds causal graphs, scores possible causes, and flags failure patterns.
Swarm 357 →
357 Claude agent roles across six business layers, with portable memory, enforced cost caps, a bash policy gate, and durable checkpoints.
FintheFinder →
FintheFinder pauses research runs for human approval and audits citations deterministically. Its repository includes contracts and test files.
TechTideAI2 →
TechTideAI2 separates agent decisions, application data, and audit output. It uses explicit state transitions, execution timeouts, and adversarial evaluation tests.
BRI →
BRI exposes video analysis tools through a FastAPI MCP server, with SQLite job storage and a Streamlit interface.
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.
All related notes
I explain the methods here. TechTide AI publishes current service scope, prices, and terms.
Systems Decision Audit at TechTide AI→© 2026 Alex Cinovoj · TechTide AI · Columbus, OH