Making Agents Visible

A real interface for terminal agent work

The Terminal Problem

Agents now do real work — the terminal shows it as a wall of text

AI agents spawn sub-agents, spend context budgets, and explore paths they don't take. A raw terminal renders all of it as one opaque, un-navigable stream.

Amplifier TUI rebuilds that terminal into an interface. First, is it real?

Proof · Not a Toy

Amplifier TUI is a substantial, well-tested codebase

The ramparte/amplifier-tui package is 171 Python files and about 36,998 lines, backed by a real test suite.

Credibility earned — so what does the wall actually hide?

171

Python files in the package

36,998

Lines of Python (approx.)

40

Test files

1,454

Test functions

Three Blind Spots

The wall hides exactly the three things you most need to see

Behind the stream: who the agent delegated to, what's eating the context budget, and the paths it didn't take.

Each of the next three beats makes one of these visible.

Who it delegated toParent→child sub-agent spawns are invisible in raw text.
What's eating the budgetContext-window consumption has no breakdown.
The paths not takenBranches and alternatives leave no trace.

Mechanism · Delegation

A first-class delegation tree tracks every sub-agent spawn

The agent tracker records when the AI spawns sub-agents via the delegate or task tool, building a tree of parent→child relationships with status and timing.

First blind spot resolved. Next: where does the budget go?

agent_tracker.pyBuilds the parent→child delegation tree.
agent_tree_panel.pyRenders the live tree with status and timing.
AGENT_DELEGATIONA dedicated block type in block_model.py.

Mechanism · Budget

A context-window profiler breaks the budget into categories

The profiler gives a categorized breakdown of context usage with stacked-bar visualization, per-message detail, sparkline history, and top-consumer identification.

Second blind spot resolved. Next: the paths not taken.

Five categoriessystem · conversation · tool · injected · meta.
Stacked bars + sparklineLive breakdown with history over time.
Top consumersNames exactly what's eating the budget.

Mechanism · Alternatives

Branches and A/B compare surface the paths not taken

Named conversation branches support fork / switch / compare, and model A/B comparison records both responses with an explicit picked choice.

Third blind spot resolved. Now — the kicker.

branch_manager.pyFork / switch / compare conversation branches.
compare_manager.pyA/B compare: response_a vs response_b, an explicit picked field.
/fork · /branch · /compareAlternatives become navigable, not lost.

The Kicker

The tool that makes AI work visible was itself built almost entirely by AI

274 of 284 commits — about 96.5% — carry the Amplifier generation attribution trailer. The interface exists because AI could build it.

Which points at the pattern worth keeping.

274 / 284

Commits carrying the Amplifier generation trailer (~96.5%)

285

"Generated with [Amplifier]" trailers

292

"Co-Authored-By: Amplifier" trailers

The Pattern

When agents do the work, the terminal has to become an interface

Visibility into delegation, budget, and alternatives is the new baseline — not a nicety. Amplifier TUI backs it with 128 slash commands, 11 themes, and web, desktop, and tmux-cockpit frontends.

Sources

Research Methodology

Feature status: Implemented

Repo: ramparte/amplifier-tui · Data as of: latest commit 2026-04-23 (c967de0)

Commands run:

Line/file counts: find . -name '*.py' | wc -l and find . -name '*.py' -exec wc -l {} + | tail -1 → 171 files, 36,998 LOC
Tests: find tests -name '*.py' | wc -l; grep -rhcE 'def test_' tests/ | paste -sd+ | bc → 40 files, 1,454 test functions
AI attribution: per-commit git log -1 --format='%B' grep for the generation trailer; git rev-list --count HEAD → 274 / 284
Trailers: git log --format='%b' | grep -iE 'co-authored-by|generated with' | sort | uniq -c → 285 / 292
Slash commands: grep -rhoE '"/[a-z][a-z0-9-]*"' --include='*.py' . | sort -u | wc -l → 128
Themes: grep -cE '^\s+"[a-z].*":' theme.py → 11
Features: head of agent_tracker.py, context_profiler.py, branch_manager.py, compare_manager.py, block_model.py

Gaps: The ~96.5% figure counts the AI generation trailer's presence on commits, not that zero human review occurred — Sam Schillace is the git author on nearly all commits. Line/file counts reflect the local checkout as of the 2026-04-23 commit and may differ from any tagged release.

Primary contributors: Sam Schillace (283 commits), Michael J. Jabbour (1 commit); Amplifier (AI) co-author on 274 of 284 commits.