Baseline evidence pack
Run one fixed task before adding scaffolding and separate model, context, state, control, and feedback failures.
Output: Baseline receipt, task contract, and first-intervention decision.
For software engineers, technical leads, and developer-platform owners
Harness Engineering
Turn a coding agent into a repository system that can understand, verify, resume, and hand off real work.
Eleven field modules retrofit one brownfield application with task contracts, repository knowledge, isolated environments, durable state, agent-readable feedback, mechanical controls, verification, recovery, and measured scaling decisions.
Open module 00- Output
- A measured repo harness with instructions, state, isolation, mechanical boundaries, end-to-end verification, run receipts, handoff, and a documented go or no-go decision.
- Effort
- 25 to 30 hours including the 5 to 7 hour capstone
- Before you start
- Comfort with Git, a terminal, TypeScript or another typed language, automated tests, and local application debugging. All labs use synthetic fixtures and sandbox environments.
Foundation → lab → project
The ordered field path.
Each lesson produces an artifact, tests it against observable criteria, then carries it into the next decision. Enter where your workflow is stuck or follow the complete sequence.
- 00Diagnose the harness gap: measure before you scaffoldRun one bounded coding task in an unmodified repository, preserve the evidence, and classify each failure before adding instructions, agents, or automation.Foundation · 75 min
- 01Write an executable task contract: give done a truth valueTurn a feature request into a versioned contract with scope, observable behavior, negative cases, permissions, stop conditions, and evidence requirements.Lab · 100 min
- 02Make the repository the system of record: give agents a mapMove durable engineering knowledge into owned, indexed repository artifacts and keep the entry file small enough to guide rather than crowd the task.Lab · 110 min
- 03Bootstrap and isolate the environment: make every run start from truthTurn undocumented local setup into a deterministic startup ritual with health checks, fixture data, worktree-safe resources, bounded cleanup, and actionable failure output.Lab · 120 min
- 04Persist state and hand off cleanly: survive a fresh contextExternalize feature status, progress, evidence, blockers, and next action so a new session can continue from repository truth instead of reconstructing a chat.Lab · 115 min
- 05Shape tools for agent-readable feedback: make the next move obviousReplace noisy commands and opaque browser failures with narrow tools that return stable codes, scoped evidence, and one safe repair path.Lab · 120 min
- 06Enforce boundaries mechanically: turn guidance into executable rulesConvert recurring architecture, permission, secret, and change-scope expectations into fast checks that fail with a repair path.Lab · 130 min
- 07Build the verification loop: grade behavior, evidence, and restraintTurn the task contract into a layered evaluator that checks the diff, real user path, failure cases, evidence, and forbidden effects before completion.Lab · 150 min
- 08Observe, budget, and recover: operate the loop when nobody is watchingInstrument state, tools, checks, cost, progress, and effects so stalled or interrupted work stops safely and resumes from an authoritative checkpoint.Lab · 145 min
- 09Scale with work graphs, then control entropyModel dependencies and ownership before parallel work, measure coordination cost, and keep generated knowledge, rules, and repository structure from drifting.Project · 165 min
- 10Capstone: retrofit a brownfield repository and defend the autonomy decisionApply the complete harness to a holdout feature, reproduce the evidence from a clean environment, and keep only layers that earn measurable value.Project · 5 to 7 hours
Hands-on projects
Build evidence of operating ability.
Agent-readable repository
Build the knowledge map, deterministic initializer, isolated worktree resources, and fresh-session handoff.
Output: Repository map, startup receipt, feature ledger, and clean handoff.
Verification control plane
Expose runtime feedback, enforce architecture and permissions, grade the real user path, and recover bounded failures.
Output: Mechanical rules, eval suite, trace, budgets, and recovery runbook.
Brownfield harness capstone
Use the full harness on a holdout feature and remove any layer that fails to produce measured value.
Output: Repo harness kit, Harness Card, before-and-after report, and go or no-go decision.
Tools, field guides, and services
Continue through the work with a clear next step.
Field guide
Harness Engineering
The runtime, control, observability, and recovery layer around an agent loop.
Open resourceField guideSpec-Driven Agentic Development
Plans and acceptance criteria as durable, reviewable engineering artifacts.
Open resourceField guideSandboxed Autonomy
Worktrees, ephemeral environments, and blast-radius control for long-running agents.
Open resourcePlaybookAwesome Harness Engineering
A maintained source radar for foundations, evals, safe autonomy, and reference harnesses.
Open resourceWhen a local harness meets a real codebase
Bring the receipt, the failed case, and the control you are unsure about.
Tenten can review repository legibility, permissions, evaluator coverage, worktree isolation, recovery, and rollout evidence before your team increases agent autonomy.